argumentation-based negotiation

Argumentation-Based Negotiation

Iyad Rahwan1, Sarvapali D. Ramchurn2, Nicholas R. Jennings2

Peter McBurney3, Simon Parsons4, Liz Sonenberg1

1 Dept. of Information SystemUniversity of Melbourne, Parkville 3010, Australia

[email protected]

[email protected]

2 School of Electronics & Computer ScienceUniversity of Southampton, Southampton, SO17 1BJ, UK

{sdr01r,nrj}@ecs.soton.ac.uk3 Dept. of Computer Science

University of Liverpool, Liverpool, L69 7ZF, [email protected]

4 Dept. of Computer & Information Science, Brooklyn CollegeCity University of New York, Brooklyn NY 11210, USA

[email protected]

2-December-2003

AbstractNegotiation is essential in settings where autonomous agents have con-flicting interests and a desire to cooperate. For this reason, mechanismsin which agents exchange potential agreements according to various rulesof interaction have become very popular in recent years as evident, forexample, in the auction and mechanism design community. However, agrowing body of research is now emerging which points out limitationsin such mechanisms and advocates the idea that agents can increase thelikelihood and quality of an agreement by exchanging arguments which in-fluence each others’ states. This community further argues that argumentexchange is sometimes essential when various assumptions about agent ra-tionality cannot be satisfied. To this end, in this article, we identify themain research motivations and ambitions behind work in the field. Wethen provide a conceptual framework through which we outline the coreelements and features required by agents engaged in argumentation-basednegotiation, as well as the environment that hosts these agents. For eachof these elements, we survey and evaluate existing proposed techniques inthe literature and highlight the major challenges that need to be addressedif argument-based negotiation research is to reach its full potential.

1

Contents

1 Introduction 3

2 Approaches to Automated Negotiation 42.1 Game-theoretic Approaches to Negotiation . . . . . . . . . . . . 42.2 Heuristic-based Approaches to Negotiation . . . . . . . . . . . . . 52.3 Argumentation-based Approaches to Negotiation . . . . . . . . . 62.4 Summary . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 8

3 External Elements of ABN Frameworks 93.1 Communication Language & Domain Language . . . . . . . . . . 9

3.1.1 State of the Art . . . . . . . . . . . . . . . . . . . . . . . 103.1.2 Challenges . . . . . . . . . . . . . . . . . . . . . . . . . . 12

3.2 Negotiation Protocol . . . . . . . . . . . . . . . . . . . . . . . . . 133.2.1 State of the Art . . . . . . . . . . . . . . . . . . . . . . . 143.2.2 Challenges . . . . . . . . . . . . . . . . . . . . . . . . . . 16

3.3 Information Stores . . . . . . . . . . . . . . . . . . . . . . . . . . 173.3.1 State of the Art . . . . . . . . . . . . . . . . . . . . . . . 173.3.2 Challenges . . . . . . . . . . . . . . . . . . . . . . . . . . 18

4 Elements of ABN Agents 194.1 Argument and Proposal Evaluation . . . . . . . . . . . . . . . . . 22

4.1.1 State of the Art . . . . . . . . . . . . . . . . . . . . . . . 244.1.2 Challenges . . . . . . . . . . . . . . . . . . . . . . . . . . 26

4.2 Argument and Proposal Generation . . . . . . . . . . . . . . . . 284.2.1 State of the Art . . . . . . . . . . . . . . . . . . . . . . . 284.2.2 Challenges . . . . . . . . . . . . . . . . . . . . . . . . . . 30

4.3 Argument Selection . . . . . . . . . . . . . . . . . . . . . . . . . . 304.3.1 State of the Art . . . . . . . . . . . . . . . . . . . . . . . 304.3.2 Challenges . . . . . . . . . . . . . . . . . . . . . . . . . . 31

5 Summary of the ABN State of the Art 32

6 Conclusions and Future Directions 36

7 Acknowledgements 39

2

1 Introduction

An increasing number of computer systems are being viewed in terms of multi-ple, interacting autonomous agents. This is because the multi-agent paradigmoffers a powerful set of metaphors, concepts and techniques for conceptualising,designing, implementing and verifying complex distributed systems (Jennings,2001). As a result, applications of agent technology have ranged from elec-tronic trading and distributed business process management to air-traffic andspacecraft control (Wooldridge, 2002; Parunak, 1999).

Here, an agent is viewed as an encapsulated computer system that is situatedin an environment and is capable of flexible, autonomous action in order to meetits design objectives (Jennings, 2000; Wooldridge, 1997). In almost all cases,such agents need to interact in order to fulfill their objectives or improve theirperformance. Generally speaking, different types of interaction mechanismssuit different types of environments and applications. Thus, agents might needmechanisms that facilitate information exchange (de Boer et al., 2003; Luo et al.,2002), coordination (Durfee, 1999; Moulin and Chaib-Draa, 1996) (in whichagents arrange their individual activities in a coherent manner), collaboration(Pynadath and Tambe, 2002; Panzarasa et al., 2002) (in which agents worktogether to achieve a common objective), and so on. One such type of interactionthat is gaining increasing prominence in the agent community is negotiation. Inan attempt to reconcile the definitions proposed by Jennings et al. (2001) andWalton and Krabbe (1995), we offer the following view:1

Negotiation is a form of interaction in which a group of agents,with conflicting interests and a desire to cooperate, try to come to amutually acceptable agreement on the division of scarce resources.

Automated negotiation among autonomous agents is needed when agentshave conflicting objectives and a desire to cooperate. This typically occurswhen agents have competing claims on scarce resources, not all of which can besimultaneously satisfied. The use of the word “resources” here is to be takenin the broadest possible sense. Thus, resources can be commodities, services,time, money, etc. In short, anything that is needed to achieve something.

In the multi-agent literature, various interaction and decision mechanismsfor automated negotiation have been proposed and studied. These include:game-theoretic analysis (Kraus, 2001; Rosenschein and Zlotkin, 1994; Sand-holm, 2002b); heuristic-based approaches (Faratin, 2000; Fatima et al., 2002;Kowalczyk and Bui, 2001); and argumentation-based approaches (Kraus et al.,1998; Parsons et al., 1998; Sierra et al., 1998). In this paper, we are concernedwith argumentation-based approaches. The main distinguishing feature of suchapproaches is that they allow for more sophisticated forms of interaction than

1Note that the precise definition of negotiation is not always stated explicitly in the lit-erature. However, we believe that this definition is a reasonable generalisation of both theexplicit and implicit definitions that can be found.

3

their game-theoretic and heuristic counterparts. This raises a number of re-search challenges related to both the design of the interaction environment aswell as the agents participating in that interaction.

In this article, we aim at setting up a research agenda for argumentation-based negotiation in multi-agent systems. We do so by achieving the follow-ing. First, we identify the main features of argumentation-based negotiationapproaches. We do this by discussing the characteristics of traditional ap-proaches and demonstrate why they fail in particular circumstances due totheir underlying assumptions. Second, we discuss, in detail, the essential el-ements of argumentation-based negotiation frameworks and the agents that op-erate within these frameworks. We do this by constructing a conceptual modelof argumentation-based negotiation, involving external elements (namely, thecommunication and domain languages, the negotiation protocol, and the in-formation stores) and agent-internal elements (namely, the ability to evaluate,generate, and select proposals and arguments). In the course of discussing eachelement, we present an overview of existing work in the literature and identifythe major challenges and opportunities that remain unaddressed.

This article is organised as follows. In the next section, we briefly review thedifferent approaches to automated negotiation and outline the contexts in whichwe believe argumentation-based approaches would be most useful. In section 3,we describe, in detail, the elements of an argumentation-based framework thatare external to the agents, namely the communication and domain languages,the negotiation protocol, and various information stores. In section 4, we moveto discussing the various internal elements and functionalities necessary to en-able an agent to conduct argumentation-based negotiation. More precisely, wediscuss the processes of argument and proposal evaluation, argument and pro-posal generation, and argument selection. In section 5, we summarise the land-scape of existing frameworks. And finally, in section 6, we state conclusions andsummarise the major research challenges.

2 Approaches to Automated Negotiation

In this section, we discuss the three major classes of approaches to automatednegotiation in the multi-agent literature. Even though there may be manyways to classify existing approaches to automated negotiation, the followingclassification suits our purpose.2

2.1 Game-theoretic Approaches to Negotiation

Game theory (Osborne and Rubinstein, 1994) is a branch of economics thatstudies the strategic interactions between self-interested economic agents.3 It

2For a more comprehensive comparison between the various approaches to automatednegotiation, see (Jennings et al., 2001).

3We say “economic” agents because economics is concerned with the interaction amongpeople, organisations, etc., rather than among computational agents.

4

has its roots in the work of von Neuman and Morgenstern (1944). Recently, ithas been used extensively to study interaction between self-interested compu-tational agents (Rosenschein and Zlotkin, 1994; Sandholm, 2002b).

In game-theoretic analysis, researchers usually attempt to determine theoptimal strategy by analyzing the interaction as a game between identical par-ticipants, and seeking its equilibrium (Harsanyi, 1956; Rosenschein and Zlotkin,1994; von Stengel, 2002). The strategy determined by these methods can some-times be made to be optimal for a participant, given the game rules, the assumedpayoffs, and the goals of the participants, and assuming that the participantshave no knowledge of one another not provided by introspection. Assuming fur-ther that participants behave according to the assumptions of rational-choicetheory (Coleman, 1990), then this approach can guide the design of the inter-action mechanism itself, and thus force such agents to behave in certain ways(Varian, 1995; Conitzer and Sandholm, 2002).

However, classical game-theoretic approaches have some significant limita-tions from the computational perspective (Dash et al., 2003). Specifically, mostthese approaches assume that agents have unbounded computational resourcesand that the space of outcomes is completely known. In most realistic environ-ments, however, these assumptions fail due to the limited processing and com-munication capabilities of the information systems. Agents may be resource-constrained, altruistic, malicious, or simply badly-coded, so that participantbehaviour may not conform to the assumptions of rational choice theory.4

2.2 Heuristic-based Approaches to Negotiation

To address some of the aforementioned limitations of game-theoretic approaches,a number of heuristic approaches have emerged. Heuristics are rules of thumbthat produce good enough (rather than optimal) outcomes and are often pro-duced in contexts with more relaxed assumptions about agents’ rationality andresources. The support for particular heuristics is usually based on empiri-cal testing and evaluation (e.g., Faratin, 2000; Kraus, 2001). In general, thesemethods offer approximations to the decisions made according to game-theoreticstudies. One example of this approach is presented by Faratin, Sierra and Jen-nings in a number of papers (see Faratin, 2000; Sierra et al., 1997). In this model,various heuristic decision functions are used for evaluating and generating of-fers or proposals (i.e., potential deals) in multi-attribute negotiation (Faratinet al., 1998). A method for generating tradeoffs is also presented which aidsthe construction of alternative offers during bargaining (Faratin et al., 2002).

4A growing research area in economics that addresses some of the limitations of conven-tional models is evolutionary game theory (Samuelson, 1998), in which the assumption ofunbounded rationality is relaxed. In evolutionary models, games are played repeatedly, andstrategies are tested through a trial-and-error learning process in which players gradually dis-cover that some strategies work better than others. However, other assumptions, such as theavailability of a preference valuation function, still hold. Another attempt is the modelling of‘bounded rationality’ by explicitly capturing elements of the process of choice, such as limitedmemory, limited knowledge, approximate preferences (that ignore minor difference betweenoptions), etc. (Rubinstein, 1997).

5

Kowalczyk and Bui (2001) present a negotiation model with decision proceduresbased on distributed constraint satisfaction (Yokoo, 1998). This was later ex-tended to allow for multiple concurrent negotiations (Rahwan et al., 2002) andto accommodate fuzzy (as opposed to ‘crisp’) constraints (Kowalczyk, 2000).The idea of using fuzzy constraint satisfaction is further investigated by Luoet al. (2003). Fatima et al. (2004, 2002, 2001) study the influence of informa-tion and time constraints on the negotiation equilibrium in a particular heuristicmodel.

While heuristic methods do indeed overcome some of the shortcomings ofgame-theoretic approaches, they also have a number of disadvantages (Jenningset al., 2001). Firstly, the models often lead to outcomes that are sub-optimalbecause they adopt an approximate notion of rationality and because they do notexamine the full space of possible outcomes. And secondly, it is very difficultto predict precisely how the system and the constituent agents will behave.Consequently, the models need extensive evaluation through simulations andempirical analysis.

2.3 Argumentation-based Approaches to Negotiation

Although game theoretic and heuristic based approaches have produced sophis-ticated systems and are highly suitable for a wide range of applications, theyshare some further limitations in addition to those mentioned above.

In most game-theoretic and heuristic models, agents exchange proposals (i.e,potential agreements or potential deals). This, for example, can be a promise topurchase a good at a specified price in an English auction, a value assignment tomultiple attributes in a multi-dimensional auction (Wurman, 1999), or an alter-nate offer in a bargaining encounter (Larson and Sandholm, 2002). Agents arenot allowed to exchange any additional information other than what is expressedin the proposal itself. This can be problematic, for example, in situations whereagents have limited information about the environment, or where their rationalchoices depend on those of other agents.5

Another limitation of conventional approaches to automated negotiation isthat agent’s utilities or preferences are usually assumed to be completely charac-terised prior to the interaction. Thus an agent is assumed to have a mechanismby which it can assess and compare any two proposals. This may be easy, forexample, when the utility of the negotiation object is defined in terms of a mon-etary value, such as the charging rate of a phone call. An agent can compareproposals of two phone service providers by simply comparing how much theycharge per minute. However, in more complex negotiation situations, such astrade union negotiations, agents may well have incomplete information whichlimits this capability. Thus, agents might:

• lack some of the information relevant to making a comparison between5This is typically the case, for instance, with network goods such as fax machines or

computer operating systems. Here, the value of a fax machine to one agent depends onwhether or not other agents have fax machines.

6

two potential outcomes,

• have limited resources preventing them from acquiring such information,

• have the information, but lack the time needed to process it in order tomake the comparison,

• have inconsistent or uncertain beliefs about the environment,

• have unformed or undetermined preferences (e.g., about products new tothem), or

• have incoherent preferences.

Thus, to overcome these limitations, the process of acquiring information, re-solving uncertainties, revising preferences, etc. often takes place as part of thenegotiation process itself.

A further drawback of traditional models to automated negotiation is thatagents’ preferences over proposals are often assumed to be proper in the sensethat they reflect the true benefit the agent receives from satisfying these pref-erences. For example, an agent attempting to purchase a car might assign ahigh value to a particular brand according to its belief that this brand makessafer cars than other brands. If this belief is false, then the preferences do notproperly reflect the agent’s actual gain if it was to purchased that car.

Finally, game-theoretic and heuristic approaches assume that agents’ utilitiesor preferences are fixed. One agent cannot directly influence another agent’spreference model, or any of its internal mental attitudes (e.g., beliefs, desires,goals, etc.) that generate its preference model. A rational agent would onlymodify its preferences upon receipt of new information. Traditional automatednegotiation mechanisms do not facilitate the exchange of this information.

Against this background, argumentation-based approaches to negotiationattempt to overcome the above limitations by allowing agents to exchange ad-ditional information, or to “argue”6 about their beliefs and other mental atti-tudes during the negotiation process. In the context of negotiation, we view anargument as a piece of information that may allow an agent to (a) justify its ne-gotiation stance; or (b) influence another agent’s negotiation stance (Jenningset al., 1998).

Thus, in addition to accepting or rejecting a proposal, an agent can offer acritique of it. This can help make negotiations more efficient. By understandingwhy its counterpart cannot accept a particular deal, an agent may be in a betterposition to make an alternative offer that has a higher chance of being accept-able. In a trade union dispute, for example, an agent representing the worker’s

6In this survey, we do not treat the topic of argumentation based on deafeasible or non-monotonic reasoning as discussed, for example, by Prakken and Vreeswijk (2002); Vreeswijk(1997); Chesnevar et al. (2000); Dung (1995); Loui (1987). Our focus here is on the generalcharacteristics of argumentation in negotiation models for multi-agent systems. One may useeither of the above argumentation systems as a basis for an ABN system.

7

union might refuse an offer for a modified pension plan made by the organisa-tion’s management agent (Sycara, 1985, 1992). As a response, the managementagent might offer a different pension plan. If the union agent had been able toexplain that the problem with the initial offer was not with its pension plan butrather that it did not include reduced working hours, the management agentwould not have bothered exploring different pension plans. Instead, the man-agement agent would have concentrated on finding an arrangement for workloadreduction.

Another type of information that can be exchanged is a justification of aproposal, stating why an agent made such a proposal or why the counterpartshould accept it. This may make it possible to change the other agent’s regionof acceptability7 (Jennings et al., 1998), or the nature of the negotiation spaceitself. For example, an employee negotiating a salary raise might propose a bigincrease that gets rejected by the manager. After the employee justifies theproposal by denoting her significant achievements during the year, the managermight accept. Agents may also exchange information that results in changingthe negotiation object itself, by introducing new attributes (or dimensions) tothe negotiation object. For example, the manager might modify the negotiationobject such that the negotiation involves not only the salary amount, but alsothe number of working hours. In this way, the manager might be able to offerreduced working hours instead of a salary increase.

An agent might also make a threat or promise a reward in order to exertsome pressure on its counterpart to accept a proposal. For example, a managerrequesting a project to be completed by a short deadline might promise a salaryraise (or threaten to fire the employees) in order to entice them to allocate moretime to working on that particular project.8

2.4 Summary

From the discussion it should be clear that there is no universal approach toautomated negotiation that suits every problem domain. Rather, there is aset of approaches, each based on different assumptions about the environmentand the agents involved in the interaction. The particular class of approachesthat we will focus on in this paper, often referred to as argumentation-basednegotiation (ABN) frameworks, is gaining increasing popularity for its potentialability to overcome the limitations of more conventional approaches to auto-mated negotiation. However, such models are typically more complex thantheir game-theoretic and heuristic counterparts.

Against this background, the aim of this analytical survey is to identify themain components of an abstract framework for ABN and discuss the different

7The region of acceptability may be defined as the complete set of outcomes the agent iswilling to accept.

8Promises and threats are also captured in evolutionary game theoretic models (Samuelson,1998). For example, by punishing noncooperative moves by its opponent, an agent sends anindirect threat for future iterations of the game. However, such threats and rewards spanover multiple, complete iterations of the same encounter, rather than being part of a singleencounter.

8

attempts to realise these components. While doing so, we hightlight the majorchallenges encountered in the field.

3 External Elements of ABN Frameworks

At this time, there is no agreed upon approach to characterising all negotiationframeworks. However, we believe it is instructive to develop such a frameworkso that the essential components that are needed to conduct automated nego-tiation, and consequently their associated challenges, can be clearly identified.In this section, we outline those elements that we consider are essential in thedesign of an ABN framework in particular. By developing an understanding ofwhat an ABN framework is expected to contain, we are in a better position tounderstand and analyse existing models that have been proposed in the litera-ture. Moreover, this nomenclature enables us to identify the ABN landscape andthe main open research questions in the field. In the course of the discussion,we outline some of the major characteristics that differentiate ABN frameworksfrom other, non-argumentation based, approaches to automated negotiation.

Abstractly, a negotiation framework can be viewed in terms of its negotiatingagents (with their internal motivations, decision mechanisms, knowledge-bases,etc.) and the environment in which these agents interact (with its rules of in-teraction, communication language, and information stores).9 In the remainderof this section, we discuss the main elements that define an ABN framework. Inparticular, we focus on the elements external to the agent (i.e., those elementsthat define the environment in which the ABN agents operate and interact).We leave the discussion of the internal features of ABN agents to section 4.

3.1 Communication Language & Domain Language

Negotiation is, by definition, a form of interaction between agents. Therefore, anegotiation framework requires a language that facilitates such communication(Labrou et al., 1999). Elements of the communication language are usually re-ferred to as locutions or utterances or speech acts (Searle, 1969; Traum, 1999).Traditional automated negotiation mechanisms normally include the basic lo-cutions such as propose for making proposals, accept for accepting proposals,and reject for rejecting proposals.

In addition to the communication language, agents often need a common do-main language for referring to concepts of the environment, the different agents,

9There are other ways in which a negotiation framework can be viewed abstractly , suchas those presented by Bartolini et al. (2002) and Wurman et al. (2001), which view auctionframeworks in terms of the rules that parametrise them. However, since these frameworksfocus on auction mechanisms, they mainly address the external rules of interaction, and donot address issues such as commitments and preference modification. We believe our model ismore suitable for the task at hand because it marks out the features peculiar to argumentationbased approaches.

9

time, proposals, and so on.10 When a statement in the domain language is ex-changed between agents, it is given particular meaning by the communicationlanguage utterance that encapsulates it. For example, in the framework pre-sented by Sierra et al. (1998), the locution offer(a, b,Price = $200 ∧ Item =palm130, t1), means that agent a proposes to agent b, at time t1 the sale ofitem palm130 for the price of $200. On the other hand, the reject locutiongives the same content a different meaning. The locution reject(b, a,Price =$200∧ Item = palm130, t2) means that agent b rejects such a proposal made byagent a.

In ABN frameworks, agents need richer communication and domain lan-guages to be able to exchange meta-level information (i.e., information otherthan that describing outcomes). Therefore, a major distinguishing factor ofABN frameworks is in the type of information that can be expressed and ex-changed between agents, and consequently, in the specifications of the agentsthat generate and evaluate this information. Table 1 shows the main distin-guishing features between ABN and non-ABN frameworks as they relate to thecommunication and domain languages.

Non-ABN Frameworks ABN FrameworksDomain Lan-guage

Expresses proposals only(e.g., by describing productsavailable for sale).

Expresses proposals as wellas meta-information aboutthe world, agent’s beliefs,preferences, goals, etc.

CommunicationLanguage

Locutions allow agents topass call for bids, propos-als, acceptance and rejec-tion, etc.

In addition, locutionsallow agents to passmeta-information eitherseparately or in conjunctionwith other locutions.

Table 1: Differences between ABN and Non-ABN w.r.t Domain and Communi-cation Languages

3.1.1 State of the Art

In existing ABN frameworks, various domain and communication languageshave been proposed. They range from those designed as simplistic domainspecific languages to more complex languages grounded in rich logical modelsof agency.

In multi-agent systems, two major proposals for agent communication lan-guages have been advanced, namely the Knowledge Query and ManipulationLanguage (KQML) (Mayfield et al., 1996) and the Foundation for Intelligent

10Note that this language may be different from the language used internally by an agent.In such cases, the agent needs to perform some type of translation into the common languagein order for communication to work (Sierra et al., 1998).

10

Physical Agents’ Agent Communication Language (FIPA ACL) (FIPA, 2001).FIPA ACL, for example, offers 22 locutions. The contents of the messages canbe in any domain language. The locution inform(a, b, ϕ, lan), for example, al-lows agent a to inform another agent b of statement ϕ which is in languagelan. Other locutions exist allowing agents to express proposals for action, ac-ceptance and rejection of proposals, make various queries about time and place,and so on. FIPA ACL has been given semantics in the form of pre- and post-conditions of each locution. This semantics is based on speech act theory, dueto a philosopher of language, John Austin (Austin, 1962) and his student JohnSearle (Searle, 1969), in which a locution is seen as an action that affects theworld in some way.

While FIPA ACL offers the benefits of being a more or less standard agentcommunication language, it fails to capture all utterances needed in a negoti-ation interaction. For example, FIPA ACL does not have locutions expressingthe desire to enter or leave a negotiation interaction, to provide an explicit cri-tique to a proposal or to request an argument for a claim. While such locutionsmay be constructed by injecting particular domain language statements withinlocutions similar to those of FIPA ACL, the semantics of these statements falloutside the boundaries of the communication language. Consider the followinglocution from the framework presented by Kraus et al. (1998):

Request(j, i,Do(i, α),Do(i, α) → Do(j, β))

In this locution, agent j requests that agent i performs action α and supportsthat request with an argument stating that if i accepts, j will perform action βin return. For the locution to properly express a promise, action β must actuallybe desired by agent i. If, on the contrary, β is undesirable to i, the same locutionbecomes a threat, and might deter i from executing α. The locution Request,however, does not include information that conveys this distinction.

In order to deal with the above problem, ABN framework designers oftenchoose to provide their own negotiation-specific locutions which hold, withinthem, the appropriate semantics of the message. For example, Sierra et al.(1998) and Ramchurn et al. (2003b) provide explicit locutions for expressingthreats and rewards (e.g., threaten(i, j, α, β) and promise(i, j, α, β)).

Having discussed some issues relating to the communication languages inABN, let us now discuss the domain languages. In negotiation, the domainlanguage must, at least, be capable of expressing the object of negotiation. InSierra et al.’s model, the domain language can express variables representingnegotiation issues (or attributes), constants representing values for the negoti-ation issues (including a special constant ‘?’ denoting the absence of value), aswell as equality and conjunction. This enables them to express full or partialmultiple-attribute proposals. For example the sentence

(Price = £10) ∧ (Quality = high) ∧ (Penalty =?)

expresses a proposal to agree on a high-quality product or service for the priceof £10, and with a cancellation penalty yet to be agreed upon. There is also a

11

meta-language for explicitly expressing preferences. For example, the statementPref([Price = £10], [Price = £20]) expresses the fact that an agent prefers aprice of £10 to £20.

In addition, ABN frameworks may need some way to express plans andresources needed for different plans. This is because agents participating innegotiation may be doing so in order to obtain resources needed for executingtheir plans. This means that an agent may be able to inform another agentof (parts of) its plans in order to justify its request for particular resources.Sadri et al. (2002), for example, express plans using the plan(.) predicate. Thefollowing formula:

plan(〈hit(nail), hang(picture)〉, {picture,nail, hammer})

denotes a plan (or intention) to hit a nail and hang a picture. The resourcesthis plan requires are a picture, a nail and a hammer.

Some ABN frameworks also explicitly express information about agents’mental attitudes. The ABN frameworks presented by Kraus et al. (1998) and byParsons et al. (1998), for example, allow an agent to represent beliefs about otheragents’ beliefs, desires, intentions, capabilities, and so on, and are based on logicsof Belief, Desire, and Intention (BDI) (Rao and Georgeff, 1995; Wooldridge,2000). An agent can use this information not only in its internal reasoningprocesses, but also in its interaction with other agents.

The usefulness of the domain language in the context of ABN becomes partic-ularly apparent when agents provide arguments for requesting certain resources,for rejecting certain requests, and so on. The richer the domain language, thericher the arguments that can be exchanged between agents. This will becomemore evident when we discuss argument generation and evaluation in the fol-lowing sections.

3.1.2 Challenges

There are a number of challenges in the design of domain and communicationlanguages for argumentation-based negotiation. First, there is a need to pro-vide rich communication languages with clear semantics. To this end, McBurneyet al. (2003) specified a set of locutions as part of a dialogue game11 for purchasenegotiation among multiple agents. The authors provided a public axiomaticsemantics to their locutions by stating each locution’s externally observablepreconditions, the possible response, and the updates to the information andcommitment stores.12 Moreover, the framework presents an operational se-mantics of the whole framework, connecting locutions with each other via the

11Dialogue games are interactions between two or more players, where each player makesa move by making some utterance in a common communication language, and according tosome pre-defined rules. Dialogue games have their roots in the philosophy of argumentation(Aristotle, 1928; Hamblin, 1970). In multi-agent systems, dialogue games have been used tospecify dialogue protocols for persuasion (Amgoud et al., 2000a), negotiation (Amgoud andParsons, 2001), and team formation (Dignum et al., 2000).

12We shall discuss information and commitment stores in more detail in section 3.3.

12

agents’ decision mechanisms. However, this framework does not cover the wholespectrum of ABN situations. For example, there are no locutions for explicitlyrequesting, providing and challenging arguments, or for supporting argumen-tation over preference criteria. Locutions facilitating argument exchange havebeen proposed in other frameworks (e.g., Sadri et al., 2001a; Torroni and Toni,2001; Sadri et al., 2002; Amgoud et al., 2000b; Amgoud and Parsons, 2001).There are opportunities for extending the model of McBurney et al. (2003) witha richer argumentation system.

Another prospect of future research is the building of common, standardiseddomain languages that agent designers can use in order to plug their agentsinto heterogeneous environments. Efforts towards semantic and syntactic in-teroperability in domain languages and ontologies, such as the DARPA AgentMarkup Language (Hendler and McGuinness, 2000; McGuinness, 2001) and theW3C Web Ontology Language (OWL) (McGuinness and van Harmelen, 2003)are particularly relevant. There is a need for exploring the suitability of thesedomain languages for supporting ABN and understanding how arguments canbe expressed and exchanged.

3.2 Negotiation Protocol

Given a communication and domain language, a negotiation framework shouldalso specify a negotiation protocol in order to constrain the use of the language.Here we view a protocol as a formal set of conventions governing the interactionamong participants (Jennings et al., 2001). This includes the interaction protocolas well as other rules of the dialogue.

The interaction protocol specifies, at each stage of the negotiation process,who is allowed to say what. For example, after one agent makes a proposal, theother agent may be able to accept it, reject it or criticise it, but might not beallowed to ignore it by making a counterproposal. The protocol might be basedsolely on the last utterance made, or might depend on a more complex historyof messages between agents.

The other rules that form part of the negotiation protocol may address thefollowing issues (Jennings et al., 2001; Esteva et al., 2001):

- Rules for admission: specify when an agent can participate in a nego-tiation dialogue and under what conditions.

- Rules for participant withdrawal: specify when a participant maywithdraw from the negotiation.

- Termination rules: specify when an encounter must end (e.g. if oneagent utters an acceptance locution).

- Rules for proposal validity: specify when a proposal is compliant withsome conditions (e.g., an agent may not be allowed to make a proposalthat has already been rejected).

13

- Rules for outcome determination: specify the outcome of the interac-tion. In an auction-based framework, this would involve determining thewinning bid(s) (Sandholm, 2002a). In argumentation-based frameworks,these rules might enforce some outcome based on the underlying theoryof argumentation (e.g., if an agent cannot construct an argument againsta request, it accepts it (Parsons et al., 1998)).

- Commitment rules: specify how agents’ commitments should be man-aged, whether and when an agent can withdraw a commitment madepreviously in the dialogue, how inconsistencies between an utterance anda previous commitment are accounted for, and so on.

In ABN, the negotiation protocol is usually more complex than those innon-ABN. By “more complex”, we mean that the protocol may involve a largernumber of locutions, and a larger number of rules. This leads to computationalcomplexity arising from processes such as checking the locutions for conformancewith the protocol given the history of locutions.


With respect to the interaction protocol, a variety of trends can be found in theABN literature. Interaction protocols can be either specified in an explicit ac-cessible format, or be only implicit and hardwired into the agents’ specification.

Explicit specification of interaction protocols may be done using finite statemachines (e.g., Sierra et al., 1998; Parsons et al., 1998). While this approach maybe useful when the interaction involves a limited number of permitted locutions,it becomes harder to specify and understand when the number of locutions andtheir interactions increases significantly. This is particularly problematic whendifferent agent designers need to look up the interaction protocol specificationto guide their agents’ design and implementation. In such cases, other forms ofprotocol specification may be more suitable.

Another way of expressing interaction protocols explicitly is using dialoguegames (e.g., as in Amgoud et al., 2000b; Amgoud and Parsons, 2001; McBur-ney et al., 2003). As mentioned above, dialogue games have the advantage ofproviding clear and precise semantics of the dialogues, by stating the pre- andpost-conditions of each locution as well as its effects on agents’ commitments.The following is the specification of a locution from the protocol presented byMcBurney et al. (2003). This locution allows a seller (or advisor) agent toannounce that it (or another seller) is willing to sell a particular option:13

Locution: willing to sell(P1, T, P2, V ), where P1 is either anadvisor or a seller, T is the set of participants, P2 is a seller andV is a set of sales options.

Preconditions: Some participant P3 must have previously uttereda locution seek info(P3, S, p) where P1 ∈ S (the set of sellers),and the options in V satisfy constraint p.

13We leave the discussion of “information stores” and “commitment stores” to section 3.3.

14

Meaning: The speaker P1 indicates to audience T that agent P2 iswilling to supply the finite set V = {a, b, . . .} of purchase-optionsto any buyer in set T . Each of these options satisfy constraint puttered as part of the prior seek info(.) locution.

Response: None required.Information Store Updates: For each a ∈ V , the 3-tuple (T, P2, a)

is inserted into IS(P1), the information store for agent P1.Commitment Store Updates: No effects.

One advantage of dialogue game protocols is that they have public axiomaticsemantics. This is because they refer only to observable pre-conditions andeffects, rather than to the agents’ internal mental attitudes. This makes iteasier to verify whether agents are conforming to the protocol.

Other frameworks implicitly hardwire the interaction protocol in the agents’internal specification (e.g., Kraus et al., 1998; Sadri et al., 2001a; Torroni andToni, 2001; Sadri et al., 2001b, 2002). In these frameworks, the interaction pro-tocol is specified using logical constraints expressed in the form of if-then rules.Since these frameworks describe a logic-based approach to agent specification(Kraus et al. (1998) implement their agents using Logic Programs, while Sadriet al. (2001b) use Abductive Logic Programs), the protocol rules are coded aspart of the agent’s program. These rules take the form P (t)∧C(t) ⇒ P ′(t+1),meaning that if the agent received performative (i.e. locution) P at time t, andcondition C was satisfied at that time, then the agent must use the performativeP ′ at the next time point. The condition C describes the rationality precon-dition in the agent’s mental state. For example, one rule might state that ifan agent received a performative which includes a request for a resource andit does not have that resource, then it must refuse the request. Note that thisconstitutes a private semantics of the protocol, and is hence harder to enforceby an external regulator.

The termination rules in negotiation protocols specified as finite state ma-chines are defined as a set of links to a final state. This is usually the case whenone agent utters a withdraw(.) or an accept(.) locution. In the framework ofMcBurney et al. (2003), a rule specifies that the dialogue ends after an agentutters the locution withdraw dialogue(.) causing either no remaining sellers orno remaining buyers in the dialogue. In some frameworks, however, no termi-nation rules have been defined, and hence the dialogue remains open even afteragreement or failure.

In relation to outcome determination rules, some frameworks determine out-comes based on the logical structure of interacting arguments. For example, inthe frameworks of Parsons et al. (1998) and Amgoud et al. (2000b), a rule spec-ifies that an agent must accept a request if it fails to produce an argumentagainst that request. A similar case occurs when agents argue about their be-liefs — an agent must accept a proposition if it fails to provide an argument forthe negation of the proposition. In this sense, outcome determination is implicitin the underlying argumentation logic. In other frameworks, such as those ofKraus et al. (1998); Ramchurn et al. (2003b), outcomes are reached through

15

uttering a specific locution explicitly (e.g., by uttering accept(.)). Agents mayutter such a locution based on some internal utility evaluation.

We shall leave the discussion of commitment rules to section 3.3, where wediscuss commitment stores.

3.2.2 Challenges

Protocols for ABN share the challenges faced in the design of argumentationprotocols in general. For example, there is a need for qualities such as fairness,clarity of the underlying argumentation theory, discouragement of disruption byparticipants, rule consistency, and so on.14

One particularly important property is that of termination. To this end,some rules for preventing certain causes of infinite dialogues have been proposed.For example, the protocol of Amgoud and Parsons (2001) does not allow agentsto repeat the exact same locutions over and over again. The intuition is thatthis would prevent the agent from, say, repeating the same question over andover again. In subsequent papers, the authors present further analysis of theoutcomes of various argumentation-based dialogues (Parsons et al., 2002, 2003).

Torroni (2002) study termination and success in an ABN framework pre-sented earlier (see Sadri et al., 2001b). Since the ABN framework is groundedin an operationally defined agent architecture based on abductive logic pro-gramming, it has been possible to study some properties by referring to themachinery of abduction. In particular, the author determined an upper limitto the maximum length of a dialogue, measured in the number of exchangedmessages. Since these results are strongly dependant on the underlying logicalsystem, it is not clear whether these results can be generalised to a variety ofprotocols without regard to the internal agent architecture.

Another important desired property in ABN protocols is that of guaranteedsuccess. Wooldridge and Parsons (2000) investigate the conditions under whichparticular logic-based negotiation protocols terminate with agreement. Theyprovide results showing the complexity of solving this problem with negotiationframeworks using different domain languages. Most interestingly, they showthat the problem of determining whether a given protocol can be guaranteedto succeed, when used with a FIPA-like communication language, is provablyintractable.

An important problem related to interaction protocols in general is that ofconformance checking. This problem is concerned with answering the questionof whether a particular utterance is acceptable, given the history and contextof interaction. Conformance checking is one of the sources of complexity indialogue systems; however, to date, it has received little attention in the ABNliterature. Recently, Huget and Wooldridge (2003) investigated applying modelchecking techniques to this problem.

Another avenue of future research is in the design of admission rules in nego-tiation protocols. While some frameworks (e.g., McBurney et al., 2003) require

14For a more elaborate discussion of the properties desired in argumentation protocols, referto (McBurney et al., 2002).

16

that agents explicitly request to enter a negotiation dialogue, to our knowledge,no ABN framework includes external rules that govern admission to the negoti-ation dialogue. One may envisage situations where only agents with particularcredentials, such as reputation or performance history, may be admitted to anegotiation. More work needs to be done on investigating the effect of differentadmission rules on the outcome of negotiation. For example, a malicious agentmay attempt to disrupt the interaction among other participants, and henceshould not be admitted. Relevant work has been done in the context of agentadmission to electronic institutions (Rodriguez-Aguılar and Sierra, 2002).

3.3 Information Stores

In some ABN frameworks, there is no explicit centralised information storeavailable. Instead, agents internally keep track of past utterances (e.g., Krauset al., 1998). However, in many negotiation frameworks there is a need to keepexternally accessible information during interaction. For example, we mightneed to store the history of utterances for future reference or to store informationabout the reputation of participants (Yu and Singh, 2002; Rubiera et al., 2001).Moreover, having external information stores makes it possible to perform somekind of enforcement of protocol-related behaviours. For example, we may beable to prevent an agent from denying a promise it has previously made.


One type of information store that is common in the argumentation literature isthe commitment store.15 Commitment stores were initially conceived by Ham-blin (1970) as a way of tracking the claims made by participants in dialoguegames. Hamblin studied dialogues over beliefs, although he was at pains tostate that commitments made in dialogue games should not be construed asnecessarily representing the real beliefs of the respective participants (Hamblin,1970, p. 257). Hamblin’s notion of commitment store has been influential inlater work on dialogue games, both in philosophy and in multi-agent systems,although the notions of commitment used sometimes differ. In the work on thephilosophy of dialogue (e.g., Walton and Krabbe, 1995) the focus is on actioncommitments, i.e., promises to initiate, execute or maintain an action or courseof actions. Commitments to defend a claim if questioned, called propositionalcommitments, are viewed as special cases of such action commitments by theseauthors. In the multi-agent systems literature the concern is usually with actioncommitments, where the actions concerned are assumed to take place outsidethe agent dialogue. For example, one agent may commit to providing a specifiedproduct or service to another agent.

Note that commitment stores should not be confused with the interactionhistory, which only records the sequence of utterances during the whole inter-

15For a more detailed discussion of commitments in multi-agent dialogues, see (Maudet andChaib-draa, 2003).

17

action.16 While the latter only form a passive storage of “unprocessed” utter-ances, commitments in commitment stores have more elaborate consequences.For example, when an agent asserts a proposition p, it may not only be com-mitted to believe that p holds, but also to defending that p (if challenged), notdenying that p, giving evidence that p, and so on (Walton and Krabbe, 1995).In the multi-agent systems literature, Singh (2000) gave “social” semantics forcommitments using modal operators in branching-time logic. This semanticsis public (i.e., is based on external observations of utterances as opposed toagents’ internal mental states) and hence can be used for specifying, and check-ing for conformance with, the interaction protocols. Amgoud et al. (2002) alsopresent a social semantics of communication based on argumentation. Anotherdifference of commitment stores in comparison with interaction histories is thatcommitment stores have specific commitment rules governing the addition andremoval of statements the agent is committed to. One rule may specify, forexample, that if the agent retracted a previously asserted claim, it must alsoretract every claim based on the former via logical deduction. Another relevantconcept is that of pre-commitment proposed by Colombetti (2000). A requestpre-commits the utterer in the sense that the utterer will be committed in casethe hearer accepts the request. Commitment stores enable us to capture suchpre-commitments.

In the ABN literature, Amgoud and Parsons (2001) define for each agenta publicly accessible commitment store. Adding statements to the commit-ment store is governed by the dialogue-game rules. For example, when anagent accepts a request for action p, then p is added to its commitment store.Agents may also be allowed to retract commitments under certain conditions.In the context of purchase negotiations, McBurney et al. (2003) dealt withthe issue of retraction differently. For example, the framework involves twolocutions: agree to buy(.) and agree to sell(.), for committing to certain re-source exchanges. Instead of providing explicit locutions for retracting thesecommitments, the authors provide additional locutions: willing to buy(.) andwilling to sell(.), which are softened versions of the former locutions, how-ever, with no commitments incurred (i.e., they are free to refuse to sell or buysomething they have previously agreed upon). This way, agents may usefullyprovide information without necessarily committing to it or having to explicitlyretract it.

3.3.2 Challenges

The representation and manipulation of information stores is not a trivial task,and has significant effects on both the performance and outcomes of negotiationdialogues. In particular, information store manipulation rules have a directeffect on the types of utterances agents can make given their previous utterances(i.e., the protocol), the properties of the dialogues (e.g., termination), and thefinal outcome (e.g., the ability to change one’s mind coherently).

16Sierra et al. (1998) use the term negotiation thread, while Sadri et al. (2001b) use theterm dialogue store.

18

Some of the key questions that need to be addressed in an ABN frame-work are: Under what conditions should an agent be allowed to retract itscommitments and how would this affect the properties of dialogues? Underwhat conditions should an agent be forced to retract its commitments to main-tain consistency? While many of these questions are being investigated in themulti-agent dialogue literature in general (Maudet and Chaib-draa, 2003), thereare issues specific to negotiation dialogues. In particular, commitments to pro-viding, requesting, and exchanging resources may require different treatmentsfrom commitments in other types of dialogues, such as persuasion or informa-tion seeking. Very little work against this problem has been done in existingABN frameworks.

4 Elements of ABN Agents

In the previous section, we discussed the different elements of an ABN frame-work that are external to the participating agents. Issues such as the interactionprotocol, commitment rules, and communication languages represent the envi-ronment in which agents operate, but often these say little about how agentsare specified, or how they reason about the interaction.

Before we get into a discussion of the general features of an ABN agent, weshall describe what constitutes (at an abstract level) a basic, non-ABN-basednegotiating agent. This will allow us to clearly contrast the ABN agent fromother negotiators, making our analysis more focused. Therefore, we begin bypresenting a conceptual model of a simple negotiator in figure 1. This captures,on a very abstract level, the main components needed by an agent in orderto be capable of engaging in negotiation.17 This model is not meant to be anidealisation of all existing models in the literature, but rather a useful startingpoint for illustrating how ABN agents are different from other types of agent.

We refer to an agent involved in negotiation interactions which largely de-pend on exchanging proposals, such as auction-based and bargaining agents, asa classical negotiating agent. This agent needs to have a locution interpretationcomponent, which parses incoming messages. These locutions usually containa proposal, or an acceptance or rejection message of a previous proposal. Theymight also contain other information about the interaction, such as the identityof the sender (especially in the case of multi-party encounters). Acceptancemessages usually terminate the encounter with a deal. A proposal may bestored in a proposal database for future reference. Proposals (or rejections) feedinto a proposal evaluation and generation component, which ultimately makesa decision about whether to accept, reject or generate a counterproposal, oreven terminate the negotiation.18 This finally feeds into the locution genera-tion component which sends the response to the relevant party or parties. A

17For a more detailed discussion of the conceptual architectures for negotiating agents, referto (Ashri et al., 2003).

18Note that the way components operate is constrained by the negotiation protocol. Forexample, in English auctions, which are highly asymmetric, one agent usually only receives

19

��

��

��

��

��

��

��

��

��

��

��

� ��

��

��

�

��

��

��

��

��

��

��

��

��

��

Figure 1: Conceptual Elements of a Classical Negotiating Agent

more sophisticated classical agent may maintain a knowledge base of its mentalattitudes (such as beliefs, desires, preferences, and so on (Wooldridge, 2002)),as well as models of the environment and the negotiation counterpart(s). Thisknowledge may be used in the evaluation and generation of proposals by judg-ing the validity and worth of proposals made (for example, by verifying whetherproposals are actually feasible and do not conflict with the current observationsof the environment). Moreover, the knowledge base may be updated in thelight of new information. However, the updates that can be made are some-what limited because the only information usually available to the agent duringthe interaction is:

1. Proposals (or bids) from the counterpart or a competitor.

2. A message rejecting a proposal initially made by the agent.

3. Other observations from the environment (e.g., a manufacturing plantagent bidding for raw material may monitor customer demand changesand bid accordingly).

The agent may be able to infer certain things from this information. For ex-ample, by receiving a rejection the agent may infer that the counterpart does

bids while others only generate them. In bargaining models (e.g., Kowalczyk and Bui, 2001),however, the protocol tends to be symmetric, allowing both agents to evaluate and generateproposals.

20

not rate certain attribute/value assignments highly. Similarly, by receiving aproposal (or by observing the proposal of another competing bidder) the agentmight infer attribute values that appeal to the counterpart (or competitor),which can then guide his own bargaining or bidding strategy.19

In contrast with a classical negotiating agent, more sophisticated meta-levelinformation can be explicitly exchanged between the ABN agents (see figure2).20 This, in turn, can have a direct effect on the agent’s knowledge base.Therefore, in addition to evaluating and generating proposals, an agent capableof participating in argument-based negotiation must be equipped with mecha-nisms for evaluating arguments (and updating the mental state accordingly) andfor generating and selecting arguments. If the locution contains an argument,

��

��

��

��

��

��

��

��

��

��

��

��

��

��

� ��

��

��

�

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

Figure 2: Conceptual Elements of an ABN Agent (the dashed lined boxes rep-resent the additional components necessary for ABN agents).

an argument evaluation or interpretation mechanism is invoked which updatesthe agent’s mental state accordingly. This may involve updating the agent’smental attitudes about itself and/or about the environment and its counter-parts. Now, the agent can enter the proposal evaluation stage in the light ofthis new information. Note that at this stage, not only does the agent evaluate

19Similar issues have been investigated in the study of signalling in game-theory (Spence,1974).

20Note that the actual way in which ABN agents are designed or implemented may differfrom the above. For example, the agent might perform certain operations in a different order,or might combine or further decompose certain functionalities. Therefore, our conceptualmodel is to be taken in the abstract sense and should not be seen as a prescriptive account ofhow ABN agents must precisely look like. Instead, it provides a useful point of departure forbeginning an analysis of the generic features of these agents.

21

the most recent proposal, but it can also re-evaluate previous proposals madeby its counterparts; these proposals are stored in the proposal database. Thisis important since the agent might (intentionally or otherwise) be persuaded toaccept a proposal it has previously rejected.

As a result of evaluating proposals, the agent may generate a counterpro-posal, a rejection, or an acceptance. In addition, however, a final argument gen-eration mechanism is responsible for deciding what response to actually send tothe counterpart, and what (if any) arguments should accompany the response.For example, the proposal evaluation and generation component might decidethat a proposal is not acceptable, and the argument generation mechanism mightaccompany the rejection with a critique describing the reasons behind the re-jection. Such arguments might also be explicitly requested by the other partyor even enforced by the protocol. Note that an agent may also choose to senda stand-alone argument (i.e., not necessarily in conjunction with a proposal,acceptance or rejection).

At times, there might be a number of potential arguments that the agentcan send. For example, in order to exert pressure on a counterpart, an agentmight be able to either make a threat or present a logical argument supportingsome action. Deciding on which argument to actually send is the responsibil-ity of an argument selection mechanism. Finally, this information is given tothe locution generation mechanism which places this information in the propermessage format and utters it.

In summary, negotiating agents must, at least, be able to:

1. interpret incoming locutions

2. evaluate incoming proposals

3. generate outgoing proposals

4. generate outgoing locutions

An ABN agent needs, in addition, to be able to:

1. evaluate incoming arguments and update its mental state accordingly

2. generate candidate outgoing arguments

3. select an argument from the set of available arguments

Now that we have given an overview of the features of an ABN agent, weconsider each of these features in more detail. In the course of doing so, weevaluate the state of the art and outline major challenges and opportunities.

4.1 Argument and Proposal Evaluation

Recall that an ABN agent needs to evaluate potential agreements proposedby its counterpart(s). The agent also needs to be able to evaluate arguments

22

intended at influencing its mental state. While proposals may be evaluated morestraightforwardly through comparison with some subjective preference criteria,argument evaluation is less trivial.

Argument evaluation is a central topic in the study of argumentation, andhas been studied extensively by philosophers at least from the days of Aristotle(Aristotle, 1928; Hitchcock, 2002). In Artificial Intelligence, argument evalua-tion and comparison has been applied, for example, in internal agent delibera-tion (Kakas and Moraitis, 2003), in legal argumentation (Prakken and Sartor,2001), and in medical diagnosis (Krause et al., 1995; Fox and Parsons, 1998).

Here, however, we find it useful to distinguish between two types of consid-erations in argument evaluation:

1. Objective Considerations: An argument may be seen as a tentativeproof for some conclusion. Hence, an agent, or a set of agents, may evalu-ate an argument based on some objective convention that defines how thequality of that proof is established. This may be done, for example, byinvestigating the correctness of its inference steps, or by examining the va-lidity of its underlying assumptions. For example, Elvang-Gøransson et al.(1993b) propose a classification of arguments into acceptability classesbased on the strength of their construction. Arguments may also be evalu-ated based on their relationships with other arguments. For Dung (1995),for instance, an argument is said to be acceptable with respect to a set S ofarguments if every argument attacking it is itself attacked by an argumentfrom that set. The set S is said to be admissible if it is conflict free andall its arguments are acceptable with respect to S.

2. Subjective Considerations: Instead of applying an objective, agent-independent convention for evaluating arguments, an agent may choose toconsider its own preferences and motivations in making that judgement,or those of the intended audience. In the framework presented by Bench-Capon (2001), for example, different participants in a persuasion dialoguehave different preferences over the “values” of arguments. Argument as-sessment and comparison would then take place in accordance with thepreferences of the dialogue participants. This means that the participantsmay influence the outcome of the argument evaluation process by havingdifferent subjective preferences.

Let us now examine the usage of the above considerations in different types ofargumentation dialogues. If two agents are reasoning about what is true in theworld (i.e., if they are conducting theoretical reasoning), then it makes sense forthem to adopt an objective convention that is not influenced by their individualbiases and motivations. For example, whether it is sunny outside should not beinfluenced by whether participants want it to be sunny, but rather only by thematerial evidence available.

If, on the other hand, two participants are engaged in a dialogue for decidingwhat course of action to take (i.e., if they are conducting practical reasoning),or what division of scarce resources to agree upon, or what goals to adopt, then

23

it would make more sense for them to consider their subjective, internal motiva-tions and perceptions, as well as the objective truth about their environment.21

Even objective facts may be perceived differently by different participants, andsuch differences in perception may play a crucial role in whether or not partic-ipants are able to reach agreement. For example, a potential airline travellermay perceive a particular airline as unsafe, while the staff of the airline itselfmay consider it to be safe. Presumably such a difference in perceptions maybe resolved with recourse to objective criteria (if any can be agreed) regardingrelative crash statistics, deaths-per-mile-flown on different airlines, etc. But if,for example, potential travellers perceive a particular airline as unsafe comparedto other airlines, despite objective evidence showing the airline to be safer thanothers, this perception may inhibit them from flying the airline anyway. Themarketing team of the airline concerned, in trying to persuade potential trav-ellers to fly with it, will have to engage in dialogue with potential customers onthe basis of those customers’ subjective perceptions, even though such percep-tion may be false. For the marketers to ignore such mis-perceptions risks thedialogue terminating without the potential customers flying the airline.

In summary, agents participating in negotiation are not concerned with es-tablishing the truth per se, but rather with the satisfaction of their needs. Hence,negotiation dialogues require agents to perform argument evaluation based onobjective as well as subjective criteria.22 In other words, agents need to performargument evaluation as part of, or in relation to, proposal evaluation.


As we argued above, argument evaluation in negotiation must involve bothobjective as well as subjective considerations, and hence must involve somesubjective assessment of proposals put forward by negotiation counterparts. Inthis subsection, we show some approaches to proposal and argument evaluationin the existing ABN literature.

One approach to proposal and argument evaluation is to assume agentsare benevolent, using the following simple normative rule: If I do not need aresource, I should give it away when asked. This approach can be found in anumber of frameworks (e.g., Parsons et al., 1998; Amgoud et al., 2000b; Sadriet al., 2001b).

Consider the following example from Parsons et al. (1998). An agent a in-tending to hang a picture would produce, after executing its planning procedure,intentions to acquire a nail, a hammer and a picture. Interactions with otheragents are only motivated in case the agent is not able to fulfill its intentionson its own. Suppose the agent does not have a nail. This leads the agent toadopt a new intention (we can call that a sub-intention) to acquire a nail, which

21Refer to Rahwan et al. (2003c) for a related comparison between argumentation over goalsand beliefs.

22Note that objective argument evaluation may also take into account certain “preferences”,such as the trust the evaluator has in the agent proposing the argument. However, this remainsaimed at establishing the truth, rather than being influenced by the agent’s personal gain.

24

may be written Ia(Have(a,nail)). If a believes that another agent b has a nail,it would generate another sub-sub-intention that b gives the nail to it, writtenIa(Give(b, a,nail)). This triggers a request to be made to agent b in the formH1 ` Ia(Give(b, a,nail), where the argument H1 constitutes the sequence ofdeductive steps taken to reach the request.23 In general, agent b accepts therequest unless it has a conflict with it. There are two types of conflicts thatwould cause b to reject the request:

1. Agent b has a conflicting intention. In argumentation terms, the agentrefuses the proposal if it can build an argument that rebuts it.

2. Agent b rejects one of the elements of the argument supporting the inten-tion that denotes the request. In argumentation terms, the agent refusesthe proposal because it can build an argument that undercuts it.

We shall explain the above two cases using the same picture-hanging example.An example of the first case is if agent b rejects the proposal because it alsoneeds the nail, say to hang a mirror; i.e., it can build an argument for theintention Ib(¬Give(b, a,nail)). This argument is based on (among other things)the intention Ib(Can(b, hang(mirror))). An example of the second case is if, inthe plan supporting the intention Ia(Give(b, a,nail)), agent a made the falseassumption that b possesses a nail, written Ba(Have(b,nail)). If b actually doesnot have a nail, then it would adopt the intention of modifying that belief,i.e. Ib(¬Ba(Have(b,nail))). Agents continue through a process of argumentexchange, which may involve recursively undercutting each other’s argumentsuntil a resolution is reached.

In order for argumentation to work, agents must be able to compare ar-guments. This is needed, for example, in order to be able to reject “weak”arguments. Parsons et al. compare arguments by classifying them into accept-ability classes based on the strength of their construction (Elvang-Gøranssonet al., 1993b). If two conflicting arguments belong to the same class, the au-thors assume the agent has some capability to perform comparisons based onutility analysis. However, they do not specify how this decision procedure isactually undertaken, nor do they specify the conditions it needs to satisfy.

A similar approach is taken by Sadri et al. (2001b). This framework, how-ever, does not involve arguing about beliefs. If an agent a receives a requestfor a resource, and needs that resource for achieving some goal ga, the agentrejects the request, unless an alternative acceptable plan for achieving ga can beproduced by the counterpart, with a promise to provide any missing resourcesfor that plan to a. Agents are also assumed to have some ordering over plansthat allows them to choose between alternative plans.

In the frameworks of Parsons et al. and Sadri et al. described above, ar-gument and proposal evaluation take into account a very simplistic subjective

23Note that the argument (or plan) may not contain intentions only, but also belief anddesire formulae about the agent and its environment. For example, in the argument H1,agent a may state the assumption that it believes b has a nail, and so on.

25

rule; that is to accept any request that the agent does not currently need. Whilethis may be useful for facilitating cooperative behaviour and making sure agentspreserve their current subjective interests, it may not be suitable in open agentsystems where agents may be purely self-interested and may refuse to provideany resources without something in return.

An alternative trend in proposal and argument evaluation is to explicitly takeinto account the utility of the agent. The basic idea is that the agent wouldcalculate the expected utility in the cases where it accepts and rejects a particularproposal. And by comparing the expected utilities in these two cases (i.e., in theresulting states), the agent would be able to make a decision about whether toaccept or reject the proposal. In the framework of Kraus et al. (1998), the agentmakes a decision about whether to accept a request by evaluating three factors:(i) the Collision Flag, which fires if the requested action conflicts with one ofthe agent’s intentions; (ii) the Convincing Factor, which is a value between0 and 1 assigned to the argument using some ad hoc rule (e.g., an appeal topast promise is assigned value 1 if the agent believes it has actually made suchpromise, and assigned 0 otherwise); and (iii) the Acceptability Value, whichinvolves a numerical calculation of utility tradeoffs in the case of accepting therequest versus rejecting it. However, it is not clear, from the paper, how thesefactors are combined to produce a final decision.

Ramchurn et al. (2003b) build on the work of Kraus et al. (1998) and takeit further by factoring the trust the agent has in its counterpart when calculat-ing the expected values. The probability of moving into the proposed state iscaptured by the uncertainty expressed in the trust value. The different factorstaken into consideration are combined using fuzzy reasoning mechanisms.

Sierra et al. (1998) introduce authority as a criteria for evaluating arguments.They present an authority graph imposed by a relation over agent roles. Thisgraph can be used to specify, for each pair of agents, which agent has higherauthority. The authors also propose a way of comparing authority levels of setsof agents. This can be used to compare arguments involving statements madeby multiple agents. An argument H1 is preferred to another H2 if and onlyif the authority level of agents involved in H1 is higher than those in H2. Asan example, the authors define a conciliatory agent, which accepts appeal-to-authority arguments regardless of the content of the justification of the appeal.This means that there would be no difference between a strong appeal and aweak (or even meaningless) one. While authority seems to be a useful factorin evaluating arguments in an organisation, it seems unreasonable to rely solelyon it. There are, therefore, opportunities for combining authority with otherargument evaluation techniques described earlier.

4.1.2 Challenges

The discussion above shows that the nature of argument evaluation dependslargely on the object of negotiation and the way agents represent and updatetheir internal mental states. For example, in the framework presented by Par-sons et al. (1998), agents are able to perform some objective argumentation

26

over their beliefs about the availability of resources, the achievability of inten-tions, and so on. This allows agents to potentially modify each other’s mentalattitudes, which may influence their preferences. In frameworks such as thoseof Ramchurn et al. (2003b) and Kraus et al. (1998), on the other hand, eval-uation is based solely on the direct comparison of expected utilities. Agentsdo not influence each other’s beliefs, but rather exert pressure on each otherby exercising their ability to influence the outcomes (for example, by making apromise or a threat). In other words, an agent would not voluntarily modify itsposition as a result of correcting its perceptions of the environment, but ratherforcedly concede on its position as a result of pressure from its counterpart.Many opportunities exist for combining the objective (belief-based) and subjec-tive (value-based) approaches to argument evaluation. For example, how canwe combine the objective evaluation of the logical form of an argument with asubjective evaluation of its consequences based on utility, trust, authority, etc.?

Another challenge is that of providing unified argumentation frameworksthat facilitate negotiation dialogues involving notions of goals, beliefs, plans,etc. Rahwan et al. (2003c) argue that systems of argumentation designed forarguing about beliefs are not readily suitable for allowing for argumentation overgoals, particularly due to the different ways conflict resolution among argumentsmust be dealt with. For example, there is a difference between attacking a goalby demonstrating that it is not achievable and attacking it by demonstratingthat it is not useful.

Rahwan et al. (2003b,c) demonstrate different ways in which goals may relateto their sub-goals, their super-goals and the agent’s beliefs. This allows one tocharacterise different types of arguments that may be provided against a partic-ular goal, and how they can, if successful, affect the agents’ mental states. Forexample, the following statement represents an argument stating that the goal ofgoing to Sydney is justified by the belief that there is a conference there (writtenjustify(confInSyd, goSyd)), that this goal is achievable by buying a ticket and ar-ranging accommodation (written achieve({buyTicket, arrangeAccomm}, goSyd)),and that it is instrumental towards the more basic goal of presenting a paper,(written instr(goSyd, presentPaper)).

〈({presentPaper}, {confInSyd}, {buyTicket, arrangeAccomm}) : goSyd〉

Following are some ways of attacking the argument for the goal goSyd:

1. Attack: Present statement ¬achieve({buyTicket, arrangeAccomm}, goSyd)Here, the counterpart attacks the relation between the subgoals and thegoal in question by arguing that buying a ticket and arranging accommo-dation are not sufficient for going to Sydney (say, one also needs to booka taxi to the airport).

Effect: The statement ¬achieve({buyTicket, arrangeAccomm}, goSyd) isremoved from the knowledge base. If no alternative plan is found, the goalis deemed unachievable and must be dropped.

27

2. Attack: Present statement instr(goPerth, presentPaper)Here, the counterpart provides an alternative plan for presenting the paperwithout having to go to Sydney (say there is a similar conference in Perth).

Effect: Statement instr(goPerth, presentPaper) is added to the agent’sknowledge base. The agent then compares the proposed plan (involvinggoing to Perth) with the existing plan (involving going to Sydney), andbased on the outcome of this comparison, the goal goSyd might be dropped(with the whole plan to which it belongs).

The paper discusses other types of possible attacks, as a preliminary step tounderstanding the space of possible influences ABN agents may (or must beable to) exert in the course of dialogue.

Another attempt to unify argumentation over beliefs with argumentationover values was presented by Fox and Parsons (1998). In this framework, Foxand Parsons distinguish between belief arguments and value arguments. A valueargument represents the value of a state or condition from a particular pointof view. The authors provide a qualitative probabilistic model for flatteningarguments relating to the same conclusion, and combining arguments relatingto different conclusions to make composite conclusions. The belief-argumentflattening and combination functions satisfy the axioms of probabilistic rea-soning about beliefs, while the value-argument functions satisfy the axioms ofutility theory. An agent can then use a third pair of flattening and combina-tion functions to derive the expected value of the situation(s) resulting fromexecuting an action. This involves considering the values of resulting states aswell as the probabilities of their occurrence. Therefore, this can be seen as anargumentation-based qualitative decision theory.

4.2 Argument and Proposal Generation

Another central problem in the study of argumentation is that of argumentgeneration. This problem is concerned with generating candidate arguments24

to present to a dialogue counterpart. These arguments are usually sent in or-der to entice the counterpart to accept some proposed agreement. Hence, innegotiation, argument and proposal generation are closely related processes.


In existing ABN frameworks, proposal generation is usually made as a result ofsome utility evaluation or planning process. For example, Sierra et al. (1998) andRamchurn et al. (2003b) assume agents have a means of generating proposalsthat increase (or maximise) their utilities. For Kraus et al. (1998), Parsons et al.(1998), and Sadri et al. (2001b), an underlying planner generates a set of actionsor resources needed to achieve some intention. Agents then request the actionsor resources they cannot achieve or obtain on their own, from other agents.If they fail to obtain immediate acceptance, they may propose to perform an

24We leave the discussion of selecting the best argument to section 4.3 below.

28

action (or set of actions) or to provide resources in return for acceptance. Thismay be done by just giving away what they do not need, or by measuring theutilities they lose in the exchange.

Proposals may be accompanied by arguments. In the framework of Krauset al. (1998), for example, agents may choose to accompany proposals witharguments generated using explicit rules. By means of an illustration, whatfollows is an informal description of the threat-generation rule for agent i:

IFA request has been sent to agent j to perform action α &j rejected this request &j has goals g1 and g2 &j prefers g2 to g1 &doing α achieves ¬g1 &doing β achieves ¬g2 &i believes doing β is credible and appropriateTHENi requests α again with the following threat:if you don’t do α, I will do β

If the rule body is satisfied, the corresponding threat will become a candidateargument. The agent may generate other candidate arguments, such as promisesor appeals, using other rules.

In a similar fashion, Ramchurn et al. (2003b) provide “preconditions” foreach argument to become a candidate argument. For example, for an agentto promise to perform action α in return for β, it must believe the counterpartactually wants α to be executed, that its gain from having α executed outweighsthe cost of it performing β, and so on.

As mentioned above, the frameworks of Parsons et al. (1998), Sadri et al.(2001b) and Amgoud et al. (2000b) take a planning approach to proposal gen-eration. Arguments are in fact generated in the process of proposal generationitself. In other words, an agent justifies a request by simply stating the truthabout its needs, plans, underlying assumptions, and so on, which ultimatelycaused the need to arise. This is different from other utility-based approachesdescribed above, where agents can, in a sense, create arguments, such as threatsand rewards, by exploiting their abilities to influence the outcomes. Of course,there is nothing that directly prevents agents from combining the two.

As described earlier, Rahwan et al. (2003b) provide a characterisation ofthe types of arguments an agent can make in relation to the goal and beliefstructures of its counterpart. This provides a more fine-grained portfolio ofcandidate arguments than those of Parsons et al. (1998), Sadri et al. (2001b)and Amgoud et al. (2000b) (where only plans or promises can be put forwardas arguments).

And finally, authority could also be used in argument generation. Sierraet al. (1998), for example, define a simple authoritarian agent, which alwaysexploits its social power by threatening whenever possible.

29

4.2.2 Challenges

More work needs to be done in order to provide a unified way of generatingarguments by considering both objective and subjective criteria. Moreover,there is a need for a complete characterisation of the space of possible arguments.This is not necessarily a trivial task since in some frameworks the number ofpossible arguments may be infinite (say, if the framework allows for nestedarguments about what may happen in the future, or nested dialogues).

More work is also needed to understand the influence of different factors,such as the interaction protocol, authority, expected utility, honesty, etc. onargument generation. Specifically, how can authority be used in constructingan argument? Should an agent believe in an argument in order to present it?Can agents bluff? These are few of the questions that need to be answeredbefore a complete framework for argument generation is achieved.

4.3 Argument Selection

Related to the problem of argument generation is that of argument selection.The question of argument selection is as follows: given a number of candidatearguments an agent may utter to its counterpart, which one is the “best” argu-ment from the point of view of the speaker?

Note that argument selection may take place in conjunction with argumentgeneration. An agent need not generate all possible arguments before it makesa selection of the most suitable one. Instead, the agent may only concern itselfwith the generation of the most suitable argument itself. In other words, theagent might prune the set of candidate arguments during the process of argu-ment generation. Whether or not this is possible, of course, depends on thenature of the argumentation framework underlying the agent’s decision making.


In the work of Kraus et al. (1998), arguments are selected according to thefollowing argument strength order, with threats being the strongest arguments:

1. Appeal to prevailing practice.

2. A counter example.

3. An appeal to past promise.

4. An appeal to self-interest.

5. A promise of future reward.

6. A threat.

The intuition is that a negotiator would progress from weak arguments up tothe strongest ones. For example, there is no need to threaten the counterpartif an appeal is sufficient to persuade him/her to accept a request. The authors

30

argue that generating appeals is less costly to the persuader than threats orrewards since the latter involve possible negative side-effects.

In Ramchurn et al.’s framework (Ramchurn et al., 2003b), agents factortrust and utility in order to decide on which candidate argument to send witha request. The following example rule is provided:

Rule 1: if trust is low and utility of the proposal is high(I need to do X and I don’t trust you)then send a strong argumentRule 2: if trust is high and utility of the proposal is low(I don’t really need to do X and I trust you)then send a weak argument

In this rule, low and high are linguistic variables manipulated using fuzzy opera-tors. The stronger an argument is, the more it is likely to lessen the opponent’strust in the proponent and the more it could coerce the opponent to changeits preferences (e.g. by making a significant threat). However, this lowering oftrust results in less cooperative behaviour which, in turn, makes it harder forthe proponent to persuade the opponent to accept its future proposals. Thusstrong arguments should only be sent when the negotiation needs to take placein the shortest time possible, when the proposal has high utility for the pro-ponent or when it is known that the other partner cannot be trusted to reacheffective agreements efficiently. Otherwise weaker arguments should be used.25

In other frameworks, argument generation is based on the relationships be-tween arguments. Agents in the framework presented by Parsons et al. (1998)provide the strongest argument possible based on the acceptability classes (e.g.,a tautological argument, if possible). For Amgoud et al. (2000b), agents comparearguments based on preferential ordering over their constituent propositions ina similar manner to that in argument evaluation (i.e., based on the argumen-tation system of Dung (1995)). Finally, for Sadri et al. (2001b), agents maycompare the costs of different alternative plans to present to the counterpart.

4.3.2 Challenges

The problem of argument selection can be considered the essence of strategy inABN dialogues in general (provided the candidate arguments contain all pos-sible arguments). However, there is very little existing work on strategies inmulti-agent dialogues. Some work is emerging that investigates strategic moveselection in persuasion dialogues (Amgoud and Maudet, 2002), as well as ininquiry and information seeking dialogues (Parsons et al., 2002, 2003). Similarwork needs to be done on argumentation based negotiation dialogues in orderto provide a sound theoretical base for potential applications. Rahwan et al.(2003a) provide a preliminary, informal attempt at charactarising strategic fac-tors in negotiation dialogues. In this work, strategies depend on various factors,

25Refer to the original paper for some related empirical observations.

31

such as the agents’ goals, the interaction protocol, the agents’ capabilities, theresources available to participants, and so on.

Suitable argument selection in a negotiation context must take into accountinformation about the negotiation counterpart. To this end, in Bayesian gametheory (Osborne and Rubinstein, 1994), a counterpart is modelled by a proba-bility distribution representing the uncertainty of the first party regarding thecounterparts’ initial information and the payoffs they receive from the differentoutcomes (i.e., action profiles). This raises the opportunity to use learning tech-niques in order to find patterns in the counterpart’s behaviour and use thesefindings in future encounters with the same (or similar) counterpart(s). Sand-holm and Crites (1995), for example, apply reinforcement learning in the contextof the iterated Prisoner’s Dilemma game to allow agents to better predict thepatterns of behaviour of their opponents. Bayesian learning in less restrictednegotiation protocols has also been investigated by Zeng and Sycara (1997). InABN, more sophisticated models of the negotiation counterparts are needed, andappropriate methods of updating these models are essential for understandingthe dynamics of opponents’ strategies, preferences, beliefs, etc. This is a partic-ularly challenging task for ABN since agents may not only model the observed‘behaviour’ of one another, but also the ‘mental attitudes’ motivating that be-haviour. Another important question is whether and how such learning agentsconverge to better and quicker deals in multiple negotiation encounters.

5 Summary of the ABN State of the Art

Having analysed the various frameworks in detail in the previous sections, wenow proceed to present a high-level view of what has been achieved in the fieldof ABN as a whole. In this way, we aim to identify the areas that have beenextensively researched, the key results in those areas, and to identify those areasthat have not received sufficient attention from the research community.

In table 2, we compare the different existing frameworks in terms of theirmain characteristics.26 Specifically, the first column describes the style of argu-mentation underlying the ABN framework. This covers the informal literaturethat motivates and provides intuitive backing of the research, as well as the for-mal theories underlying the specification of the framework (e.g., decision theory,argumentation theory, dialogue games, etc.). When taken together this providesan idea of the starting point of each framework. As can be seen, frameworkssuch as those of Amgoud et al. (2000b) and Sadri et al. (2001b) start froma single-agent proof procedure and try to split it into multiple disjoint agentswhile preserving the correctness of the proof theory. In particular, Amgoudet al. (2000b) build on an existing framework for belief-based argumentation ina single agent (Amgoud and Cayrol, 1998), and view the proof theory of thatframework as a dialogue between two agents. Sadri et al. (2001b) take a similarapproach by providing a dialectical version of an abductive logic programming

26Wherever the framework in question has not addressed the particular attribute of thetable (e.g. Protocol) significantly, we note this as N/A.

32

framework (Fung and Kowalski, 1997). In contrast, frameworks such as thosepresented by McBurney et al. (2003) and Rahwan et al. (2003b) start by dis-cussing the different types of interaction patterns needed among agents, andfrom there attempt to create a dialogue system. The former approach has theadvantage of being built on solid theories that make the interaction, in a sense,more predictable. On the other hand, the latter approach tends not to spec-ify the internal decision mechanisms of the agents, and concentrates instead onstudying the general properties of the dialogue itself or the interesting types ofinfluences that agents might be able to exert on one another. We believe thatas the field of argumentation-based negotiation matures, these approaches willmeet in the middle, achieving rich interactions on the system level, as well ascomprehensive and verifiable proof theories on the detailed level.

The next column describes the protocol. It is clear that some ABN frame-works have not yet addressed the protocol definition, while in others it is themainstay of their contribution. Moreover, the frameworks can differ in theway they specify the protocols, by making them implicit or explicit, definingthem as finite state machines, as dialogue games, and so on. The third col-umn describes some of the important assumptions that each framework makes.In some frameworks, for instance, the agents must be cooperative for ABN towork. Frameworks can also vary in their assumptions about agents’ utilitiesand preferences. Finally, we have specified whether the framework has beenimplemented and, if so, what form this takes.

In table 3, we outline the various frameworks in terms of whether and howeach framework addresses the problems of argument generation, selection, andevaluation. One important observation from this is that argument selection hashad very little attention in the ABN community. This is, we believe, partlybecause effective strategies for deciding what arguments to utter are likely tobe protocol-dependent. There is still no formal theory of interaction protocolscovering all types of mechanism. It is to be expected therefore that such workwill focus first on defining the protocols and exploring their properties, ratherthan on devising strategies for participants using the protocols.

As can be seen, there is a clear contrast in the way the three main mecha-nisms are conceived by the different frameworks. We contend that this is mainlydue to the differences in the underlying style of argumentation. However, despitethese differences, their contributions are broadly complementary.

In summary, the frameworks reviewed in this article represent different pre-liminary attempts at solving parts of the puzzle by: (i) constructing genericmodels of ABN (Sierra et al., 1998); (ii) constructing limited, yet implementablesystems, and studying their applicability (Kraus et al., 1998; Sadri et al., 2001b);(iii) studying the applicability of particular logic-based argumentation frame-works to ABN (Parsons et al., 1998; Amgoud et al., 2000a); (iv) studying theproperties of different decision making components and concepts such as trustin controlled settings (Ramchurn et al., 2003b); (v) constructing specific com-plete negotiation protocols necessary for facilitating particular types of ABN(McBurney et al., 2003); and (vi) studying the different types of influences thatcan be attempted by participants in an ABN dialogue (Rahwan et al., 2003c,b).

33

Char

acte

rist

icA

rgum

enta

tion

styl

ePro

toco

lM

ain

Ass

um

ption

sIm

ple

men

tation

Auth

ors

Kra

us,Sy

cara

and

Eve

nchi

k-B

ase

don

psy

cholo

gy

ofper

suasi

on

(Karl

ins

and

Abel

son,1970)

-U

seth

reats

,re

ward

sand

appea

ls

-Im

plici

tin

agen

tsp

ecifi

ca-

tion

-A

gen

tshav

euti

lity

funct

ion

-A

gen

tssh

are

sam

earc

hit

ectu

re

-B

lock

sW

orl

dSce

nari

o(i

mple

-m

ente

d)

Sier

ra,J

enni

ngs,

Nor

iega

,an

dPar

sons

-Sam

eas

abov

e-

Fin

ite

Sta

teM

ach

ine

-A

llow

spass

ing

gen

eric

met

ain

form

ati

on

-A

rgum

entgen

erati

on,se

lect

ion

and

evalu

ati

on

are

pre

defi

ned

-E

-Insi

tuti

on

pre

sent

-N

/A

Ram

chur

n,Je

nnin

gsan

dSi

erra

-Sam

eas

abov

e-

Fra

mew

ork

base

don

(Sie

rra

etal.,1998)

-N

/A

-Sta

tes

of

the

worl

dhav

ea

know

nva

lue

-Tru

stm

odel

implici

t

-A

bst

ract

worl

d(s

imula

ted)

Par

sons

,Je

nnin

gsan

dSi

erra

-Logic

-base

d,in

spir

edby

(Elv

ang-G

øra

nss

on

etal.,1993a)

-Fin

ite

Sta

teM

ach

ine

-A

llow

spass

ing

gen

eric

met

ain

form

ati

on

-A

gen

tsare

cooper

ati

ve

-A

gen

tssh

are

sam

earc

hit

ectu

re

-N

/A

Am

goud

,M

aude

tan

dPar

sons

-C

lass

ifica

tion

ofdia

logues

base

don

(Walt

on

and

Kra

bbe,

1995)

-Logic

-base

d

-D

ialo

gue

Gam

epro

to-

col,

allow

sse

ndin

gpro

mis

es,

challen

ges

,ass

erti

ons,

and

reques

ts

-A

gen

tssh

are

sam

earc

hit

ectu

re-

Com

ple

tepre

fere

nce

sov

erknow

ledge

base

s-

Agen

tssh

are

pre

fere

nce

s

-N

/A

Sadr

i,Tor

roni

and

Ton

i-

Base

don

Abduct

ive

Logic

Pro

gra

mm

ing

-V

iew

ofdia

logue

from

(Walt

on

and

Kra

bbe,

1995)

-Im

plici

tin

agen

tsp

ecifi

ca-

tion

-A

gen

tssh

are

sam

earc

hit

ectu

re-

Agen

tsare

cooper

ati

ve

-Log

icag

ents

nego

tiat

eto

ex-

chan

geto

ols

McB

urne

y,va

nE

ijk,Par

sons

,A

mgo

ud

-V

iew

ofneg

oti

ati

on

base

don

(Wal-

ton

and

Kra

bbe,

1995)

-A

gen

tm

odel

influen

ced

by

consu

mer

model

ling

model

sin

mark

etin

g(L

ilie

net

al.,1992)

-D

ialo

gue

gam

epro

toco

lal-

low

spass

ing

pote

nti

aloffer

s,pre

fere

nce

state

men

ts

-A

gen

tsdo

not

dec

ide

as

agro

up,

goods

purc

hase

dafford

neg

oti

ati

on

-A

gen

thas

uti

lity

funct

ion

and

oth

eragen

tsm

ight

know

its

val-

uati

on

for

goods

-N

/A

Rah

wan

,So

nenb

erg,

and

Dig

num

-A

ppro

ach

influen

ced

by

inte

rest

-base

dneg

oti

ati

on

am

ong

hum

ans

(Fis

her

and

Ury

,1983)

-Sket

ched

locu

tions

-A

gen

tshav

eco

mpari

son

crit

e-ri

afo

rse

lect

ing

goals

base

don

thei

rsu

pport

-N

/A

Tab

le2:

Cha

ract

eris

tics

ofE

xtan

tA

BN

mod

els

34

Mec

han

ism

/A

uth

ors

Gen

erat

ion

Sel

ection

Eva

luat

ion

Kra

us,

Syca

ra,

and

Eve

nchi

k-

Rul

esde

term

ine

poss

ible

argu

-m

ents

from

avai

labl

epo

ol-

Rig

idor

der

amon

gar

gum

ent

clas

ses

only

(sta

rtw

ith

appe

als,

then

rew

ards

,th

enth

reat

s)

-Fir

stus

ing

rule

s,th

enta

king

the

confl

icti

ngin

tent

ions

into

ac-

coun

tSi

erra

,Jen

ning

s,N

orie

ga,

and

Par

sons

-Par

tial

lysp

ecifi

edby

role

s-

N/A

-O

nly

rule

sba

sed

onau

thor

ity

are

pres

ente

d

Ram

chur

n,Je

n-ni

ngs

and

Sier

ra-

Pos

sibl

ear

gum

ents

chos

enfr

ompo

olac

cord

ing

topr

econ

diti

ons

-Tak

ing

into

acco

unt

utili

tyof

prop

osal

and

trus

tto

choo

sear

-gu

men

tof

part

icul

arst

reng

th

-A

ccor

ding

tout

ility

gain

edan

dtr

ust

Par

sons

,Je

n-ni

ngs,

and

Sier

ra

-St

arti

ngfr

omex

isti

ngin

tent

ions

,ge

nera

tete

ntat

ive

plan

san

dpr

oofs

(arg

umen

ts)

usin

gsi

mpl

eB

DI

plan

-ni

ngru

les

-Pos

sibl

ych

oose

stro

nges

tar

-gu

men

tba

sed

onac

cept

abili

tycl

asse

s

-B

ased

onac

cept

abili

tycl

asse

s-A

ccep

ted

ifar

gum

ent

cann

otbe

logi

cally

defe

ated

Am

goud

,M

aude

t,an

dPar

sons

-N

/A-

N/A

-B

ased

onin

tera

ctio

nbe

twee

nar

gum

ents

and

pref

eren

ces

over

thei

rco

nten

tsSa

dri,

Tor

roni

and

Ton

i-

Bas

edon

rule

sN

/A-

Com

pare

plan

sw

ith

orw

itho

utex

chan

geof

reso

urce

sM

cBur

ney,

van

Eijk

,Par

sons

and

Am

goud

-N

/A-

N/A

-In

form

ally

spec

ified

Rah

wan

,So

nen-

berg

,D

ignu

m-

Bas

edon

alis

tof

poss

ible

atta

cks

ongo

als

N/A

-B

ased

onth

eeff

ect

ofup

dati

ngre

lati

onsh

ips

betw

een

goal

s,su

-pe

rgoa

ls,su

bgoa

lsan

dbe

liefs

Tab

le3:

Nec

essa

rym

echa

nism

sfo

ran

AB

Nfr

amew

ork

35

6 Conclusions and Future Directions

Argumentation is gaining increasing importance as a fundamental concept inmulti-agent interaction, mainly because it enables rational dialogue (McBur-ney, 2002) – the giving and receiving of reasons for statements – and becauseit enables richer forms of negotiation than have hitherto been possible in game-theoretic or heuristic-based models. Against this background, this article hassketched the landscape of the emerging research field of argumentation basednegotiation (ABN) and reviewed the state of the art. Specifically, we identifiedthe key advantages that argumentation adds to existing models of negotiation.Then, we provided a sketch of the different features and decision componentsneeded to facilitate ABN and used these to describe and analyse a variety ofexisting ABN frameworks. Even though the research area is still in its infancy,many important lessons have been learned through the various attempts thusfar. We now have insights about the various functionalities needed for ABN,some possible ways to achieve these functionalities, and, most importantly, ma-jor challenges and open questions in the field.

This analysis has highlighted the fact that there is clearly a need for morework on identifying different ways of implementing mechanisms for argumentevaluation, generation and selection. In particular, this involves understandinghow these processes are related to the agent’s underlying notions of rationality.In other words, there is a need for a better understanding of how agents may useobjective argument-based reasoning about the state of the world to achieve sub-jective goals, such as maximising utility, satisfying preferences, fulfilling socialnorms, and so on.

This raises fundamental questions on the relationship between argumentation-based approaches to group decision making and the game-theoretic approachesdictated by the traditional economic conception of rationality. Young (2001)argues that the traditional notion of economic rationality fails to capture theprocess by which agents form their preferences during negotiation. In Young’swords :

“. . . since, under the assumptions of the received concept of economicrationality, each player’s objectives or ends are not justifiable ratio-nally, and are set by the agent’s preferences in advance of reasoning,game theory has no way to capture or evaluate those ends. It cannever be sure it understands the complete motivational set which isdriving the individual, and in the end can only reduce the rich mix offactors which motivate and guide human beings to simple economicself-interest. Unfortunately, . . .much is lost along the way.” (Young,2001, p. 97)

Young then argues that in order to solve this problem, agents must adopt anotion of communicative rationality rather than merely strategic, instrumentalrationality. We take the position that argumentation, which allows agents tocritically evaluate their and each others’ underlying motivational attitudes dur-ing negotiation, can enable agents to realise deals not possible by following the

36

game-theoretic model of negotiation. However, the exact ‘formal’ relationshipbetween ABN and game-theoretic models of strategic decision making remainsa fertile area of research.

Another important challenge facing future research is the understanding of‘social’ aspects of ABN in societies of agents. There is some existing work oninvestigating the effects of norm adoption (Glass and Grosz, 2003; Castelfranchiet al., 1998) and social influence (Panzarasa and Jennings, 2002; Sen et al.,2002) on decision making, both from the individual and collective agent per-spectives. However, there is no generic formal theory that establishes a preciserelationship between normative social behaviour and the resulting outcomes ofcommunication. A full investigation of these issues will also need to respondto the theory of communicative action of Habermas (1984), and the associatedphilosophical issues concerning group decision-making.

Another important social aspect is ‘trust’, which plays a crucial role whenagents cannot be assumed to keep to their commitments. Trust enables theselection of the most appropriate negotiation partners and affects the way in-teraction unfolds. It also enables agents to improve the outcome in repeatedencounters. For example, in iterated game-theory, trust has proven useful inallowing agents to converge to higher-payoff Nash equilibria (Mukherjee et al.,2001). There are opportunities for using trust models in ABN, over single andrepeated encounters, in order to guide the selection of negotiation strategies. Inthe context of ABN, Ramchurn et al. (2003b) use trust in argument evaluationand generation in an ABN setting. A more elaborate model of trust could alsobe incorporated (e.g., Ramchurn et al., 2003a; Sabater and Sierra, 2002). In themore general context of argumentation, Parsons and Giorgini (2000) explore theevaluation of the strength of arguments based on the reliability of the agentsproviding these arguments.

One of the significantly unexplored topics is mediated negotiation, in whicha trusted, neutral third party assists negotiators in reconciling their views andinterests. Several multi-party negotiation systems with artificial mediators havebeen built to aid human negotiation. For example, the PERSUADER system(Sycara, 1985, 1992) is a mediating system that relies on the construction ofhierarchical goal structures for participants and attempts to reconcile these.However, from the descriptions given, it is not apparent how the framework,which mostly describes the reasoning mechanism for a single agent mediatingbetween human opponents, could be extended to (artificial) multi-agent sys-tems. In another example, the ZENO system (Gordon and Karacapilidis, 1997;Gordon et al., 1997) permits human participants to undertake joint deliberationabout urban planning decisions, and provides an intelligent assistant to a humanmediator. Major issues for these systems concern the nature of the mediator —is it a human or a software agent? — and its relationship to the negotiators —does it merely assist them in identifying new proposals, or can it enforce its willon them? What is still absent in this area is a grounded theory of mediation ina multi-agent negotiation.

Understanding the computational complexity of the ABN process is impor-tant before ABN frameworks can be used in real-world applications. In this

37

context, complexity may arise, for example, from:

- The internal mental processes of the participants (e.g., preference deter-mination, argument assessment, strategy generation, strategy selection,argument generation, argument selection, etc.).

- The assessment of uttered locutions by the agent platform to determineconformance with the protocol participants.

- The achievement of particular negotiation outcomes. (e.g. complexity ofsuccessful termination).

Some work has been done on aspects of argumentation-based negotiationcomplexity. For example, relative to the assessment of uttered illocutions andconformance to protocols, Wooldridge and Parsons (2000) study complexityissues in logic-based negotiation protocols. More recently, with regards to theachievement of dialogue outcomes, Parsons et al. (2002) study the terminationand complexity of information seeking, inquiry and persuasion dialogues.

In addition to the work reported in this survey, there have been attempts tostudy the process of argumentation-based negotiation from a higher level of ab-straction and analyse some of the convergent properties associated with certainassumptions (very much like work in game theoretic models discussed in section2.1 above). For example, Tohme (1997) views negotiation as resource allocationwith uncertainty caused by imperfect information about others. The agents ex-change messages that correspond to updates of beliefs and consequently causenew messages to be generated. Tohme shows that under certain conditions,beliefs converge in the long run, leading to successful negotiation (i.e. agree-ment). He shows, again under certain conditions, that such agreement couldtake place even though no interaction protocol has been defined. There is aneed for further studies in this direction in order to understand the general dy-namics of ABN systems without necessarily subscribing to particular internalcomputational mechanisms within the agents.

In conclusion, we see argumentation-based negotiation as an important andchallenging research area that combines the study of agent architectures, decision-making, relationships, and interaction, from computer science, economics, andorganization theory, with the study of argumentation in dialogue, logic andpsychology. Existing work has begun to address different aspects of the chal-lenge, but much remains to be done, on both the conceptual and technicallevels. Argumentation-based negotiation will enable us to build more sophis-ticated, flexible, and robust negotiating agents, capable of operating in moredynamic, uncertain, and unpredictable environments. Moreover, research intoargumentation-based negotiation has the potential to inform our understandingof strategic interaction among humans, and hence to contribute to related fields.

38

7 Acknowledgements

During this work, Iyad Rahwan was supported by a Melbourne Research Schol-arship (MRS) and a Postgraduate Overseas Research Experience Scholarship(PORES) from the University of Melbourne, and a Ph.D. top-up scholarshipfrom CSIRO Mathematical and Information Sciences, Australia. Sarvapali D.Ramchurn is supported by the FEEL project (IST-2000-26135) funded by theEU commission through the Disappearing Computer initiative. This work com-menced while the first author was a visitor to the IAM Group of the School ofElectronics and Computer Science at the University of Southampton, and theAgent ART Group of the Department of Computer Science at the University ofLiverpool, UK. This research was also partially support by NSF-REC-02-19347.

References

L. Amgoud and C. Cayrol. On the acceptability of arguments in preference-based argumentation. In Proceedings of the 14th Conference on Uncertaintyin Artificial Intelligence (UAI 1998), pages 1–7, San Francisco CA, USA,1998. Morgan Kaufmann Publishers.

L. Amgoud and N. Maudet. Strategical considerations for argumentative agents(preliminary report). In S. Benferhat and E. Giunchiglia, editors, Proceed-ings of the 9th International Workshop on Non-Monotonic Reasoning (NMR2002): Special session on Argument, Dialogue and Decision, pages 399–407,2002.

L. Amgoud, N. Maudet, and S. Parsons. Modelling dialogues using argumenta-tion. In E. Durfee, editor, Proceedings of the 4th International Conference onMulti-Agent Systems (ICMAS 1998), pages 31–38, Boston MA, USA, 2000a.IEEE Press.

L. Amgoud, N. Maudet, and S. Parsons. An argumentation-based semantics foragent communication languages. In F. V. Harmelen, editor, Proceedings ofthe European Conference on Artificial Intelligence (ECAI-2002), pages 38–42,Lyon, France, July 2002. IOS Press.

L. Amgoud and S. Parsons. Agent dialogues with conflicting preferences. In J.-J.Meyer and M. Tambe, editors, Proceedings of the 8th International Workshopon Agent Theories Architectures and Languages (ATAL 2001), pages 190–205,Seattle, 2001.

L. Amgoud, S. Parsons, and N. Maudet. Arguments, dialogue, and negotiation.In W. Horn, editor, Proceedings of the European Conference on ArtificialIntelligence (ECAI 2000), pages 338–342, Amsterdam, Netherlands, 2000b.IOS Press.

Aristotle. Topics. Clarendon, Oxford, UK, 1928. (W. D. Ross, Editor).

39

R. Ashri, I. Rahwan, and M. Luck. Architectures for negotiating agents. InV. Marik, J. Muller, and M. Pechoucek, editors, Multi-Agent Systems andApplications III: Proceedings of the 3rd International Central and EasternEuropean Conference on Multi-Agent Systems (CEEMAS 2003), volume 2691of Lecture Notes in Artificial Intelligence, pages 136–146. Springer Verlag,Berlin, Germany, 2003.

J. L. Austin. How To Do Things with Words. Oxford University Press, Oxford,UK, 1962. Originally delivered as the William James Lectures at HarvardUniversity in 1955.

C. Bartolini, C. Preist, and N. R. Jennings. Architecting for reuse: A soft-ware framework for automated negotiation. In F. Giunchiglia, J. Odell,and G. Weiss, editors, Proceedings of the 3rd International Workshop onAgent-Oriented Software Engineering (AOSE-2002), Berlin, Germany, 2002.Springer Verlag.

T. J. M. Bench-Capon. Truth and consequence: Complementing logic withvalue in legal reasoning. Information and Communications Technology Law,10(1), 2001.

C. Castelfranchi, R. Conte, and M. Paolucci. Normative reputation and thecosts of compliance. Journal of Artificial Societies and Social Simulation,1998.

C. I. Chesnevar, A. Maguitman, and R. Loui. Logical models of argument. ACMComputing Surveys, 32(4):337–383, 2000.

J. S. Coleman. Foundations of Social Theory. Harvard University Press, Cam-bridge MA, USA, 1990.

M. Colombetti. A commitment-based approach to agent speech acts and con-versations. In Proceedings of the Workshop on Agent Languages and Conver-sational Policies, pages 21–29, 2000.

V. Conitzer and T. Sandholm. Complexity of mechanism design. In A. Darwicheand N. Friedman, editors, Proceedings of the 18th Conference on Uncertaintyin Artificial Intelligence (UAI-2002), pages 103–110, August 2002.

R. K. Dash, D. C. Parkes, and N. R. Jennings. Computational mechanismdesign: A call to arms. IEEE Intelligent Systems, 18(6):40–47, 2003.

F. de Boer, R. M. van Eijk, W. van der Hoek, and J.-J. Meyer. A fully abstractmodel for the exchange of information in multi-agent systems. TheoreticalComputer Science, 290(3):1753–1773, 2003.

F. Dignum, B. Dunin-Keplicz, and R. Berbrugge. Agent theory for team forma-tion by dialogue. In C. Castelfranchi and Y. Lesperance, editors, IntelligentAgents VII: Proceedings of the 7th International Workshop on Agent Theories,Architectures, and Languages (ATAL-2000), volume 1986 of Lecture Notes inComputer Science, pages 150–166. Springer Verlag, Berlin, Germany, 2000.

40

P. M. Dung. On the acceptability of arguments and its fundamental role innonmonotonic reasoning, logic programming and n-person games. ArtificialIntelligence, 77(2):321–358, 1995.

E. H. Durfee. Practically coordinating. Artificial Intelligence Magazine, 20(1):99–116, Spring 1999.

M. Elvang-Gøransson, P. Krause, and J. Fox. Acceptability of arguments as‘logical uncertainty’. In M. Clarke, R. Kruse, and S. Moral, editors, Proceed-ings of the European Conference on Symbolic and Quantitative Approaches toReasoning and Uncertainty (ECSQARU 1993), number 747 in Lecture Notesin Computer Science, pages 85–90. Springer Verlag, Berlin, Germany, 1993a.

M. Elvang-Gøransson, P. Krause, and J. Fox. Dialectic reasoning with inconsis-tent information. In D. Heckerman and A. Mamdani, editors, Proceedings ofthe 9th Conference on Uncertainty in Artificial Intelligence, pages 114–121,Washington D. C., USA, 1993b. Morgan Kaufmann.

M. Esteva, J. A. Rodrıguez, C. Sierra, P. Garcia, and J. L. Arcos. On theformal specifications of electronic institutions. In F. Dignum and C. Sierra,editors, Agent Mediated Electronic Commerce (The European AgentLink Per-spective), volume 1991 of Lecture Notes in Artificial Intelligence, pages 126–147. Springer Verlag, Berlin, Germany, 2001.

P. Faratin. Automated Service Negotiation Between Autonomous ComputationalAgents. PhD thesis, University of London, Queen Mary and Westfield College,Department of Electronic Engineering, 2000.

P. Faratin, C. Sierra, and N. R. Jennings. Negotiation decision functions forautonomous agents. International Journal of Robotics and Autonomous Sys-tems, 24(3–4):159–182, 1998.

P. Faratin, C. Sierra, and N. R. Jennings. Using similarity criteria to maketrade-offs in automated negotiations. Artificial Intelligence, 142(2):205–237,2002.

S. Fatima, M. Wooldridge, and N. R. Jennings. Optimal negotiation strategiesfor agents with incomplete information. In J.-J. Meyer and M. Tambe, editors,Intelligent Agent series VIII: Proceedings of the 8th International Workshopon Agent Theories, Architectures, and Languages (ATAL 2001), volume 2333of Lecture Notes in Computer Science, pages 53–68. Springer Verlag, Berlin,Germany, 2001.

S. Fatima, M. Wooldridge, and N. R. Jennings. Multi-issue negotiation undertime constraints. In C. Castelfranchi and L. Johnson, editors, Proceedings ofthe 1st International Joint Conference on Autonomous Agents and MultiagentSystems (AAMAS-2002), pages 143–150, New York, USA, 2002. ACM Press.

S. Fatima, M. Wooldridge, and N. R. Jennings. An agenda based framework formulti-issues negotiation. Artificial Intelligence Journal, 152(1):1–45, 2004.

41

FIPA. Communicative Act Library Specification. Technical Report XC00037H,Foundation for Intelligent Physical Agents, 10 August 2001.

R. Fisher and W. Ury. Getting to Yes: Negotiating Agreement Without GivingIn. Penguin Books, New York, USA, 1983.

J. Fox and S. Parsons. Arguing about beliefs and actions. In A. Hunter andS. Parsons, editors, Applications of Uncertainty Formalisms, number 1455 inLecture Notes in Computer Science, pages 266–302. Springer Verlag, Berlin,Germany, 1998.

T. Fung and R. Kowalski. The iff proof procedure for abductive logic program-ming. Journal of Logic Programming, 33(1):151–165, 1997.

A. Glass and B. Grosz. Socially conscious decision-making. Autonomous Agentsand Multi-Agent Systems, 6(3):317–339, May 2003.

T. F. Gordon and N. Karacapilidis. The Zeno argumentation framework. InProceedings of the Sixth International Conference on AI and Law, pages 10–18, New York, NY, USA, 1997. ACM Press.

T. F. Gordon, N. Karacapilidis, H. Voss, and A. Zauke. Computer-mediatedcooperative spatial planning. In H. Timmermans, editor, Decision SupportSystems in Urban Planning, pages 299–309. E and F. N. Spon, 1997.

J. Habermas. The Theory of Communicative Action, Volume 1 : Reason andthe Rationalization of Society. Beacon Press, Boston MA, USA, 1984.

C. L. Hamblin. Fallacies. Methuen, London, UK, 1970.

J. C. Harsanyi. Approaches to the bargaining problem before and after thetheory of games: a critical discussion of Zeuthen’s, Hicks’, and Nash’s theories.Econometrica, 24:144–157, 1956.

J. Hendler and D. L. McGuinness. The DARPA agent markup language. IEEEIntelligent Systems, 15(6):67–73, November/December 2000.

D. Hitchcock. Aristotle’s theory of argument evaluation. In K. Boudourisand J. Poulakos, editors, The Philosophy of Communication, volume 1. IoniaPress, Athens, Greece, 2002.

M.-P. Huget and M. Wooldridge. Model checking for acl compliance verification.In Proceedings of the Workshop on Agent Communication Languages andConversation Policieis, page (to appear), 2003.

N. R. Jennings. On agent-based software engineering. Artificial Intelligence,117(2):277–296, 2000.

N. R. Jennings. An agent-based approach for building complex software systems.Communications of the ACM, 44(4):35–41, 2001.

42

N. R. Jennings, P. Faratin, A. R. Lomuscio, S. Parsons, C. Sierra, andM. Wooldridge. Automated negotiation: prospects, methods and challenges.International Journal of Group Decision and Negotiation, 10(2):199–215,2001.

N. R. Jennings, S. Parsons, P. Noriega, and C. Sierra. On argumentation-basednegotiation. In Proceedings of the International Workshop on Multi-AgentSystems, pages 1–7, Boston, USA, 1998.

A. Kakas and P. Moraitis. Argumentation based decision making for au-tonomous agents (to appear). In Proceedings of the 2nd International JointConference on Autonomous Agents and Multiagent Systems (AAMAS-2003),pages 883–890, Melbourne, Australia, 2003.

M. Karlins and H. I. Abelson. Persuasion: How Opinions and Attitudes areChanged. Lockwood, second edition, 1970.

R. Kowalczyk. On negotiation as a distributed fuzzy constraint satisfactionproblem. In Proceedings of the 3rd International Symposium on Soft Com-puting for Industry, World Automation Congress, pages 631–637, 2000.

R. Kowalczyk and V. Bui. On constraint-based reasoning in e-negotiationagents. In F. Dighum and U. Cortes, editors, Agent-Mediated Electronic Com-merce III, volume 2003 of Lecture Notes in Computer Science, pages 31–46.Springer Verlag, Berlin, Germany, 2001.

S. Kraus. Strategic Negotiation in Multi-Agent Environments. MIT Press, Cam-bridge MA, USA, 2001.

S. Kraus, K. Sycara, and A. Evenchik. Reaching agreements through argumen-tation: A logical model and implementation. Artificial Intelligence, 104(1–2):1–69, 1998.

P. Krause, S. Ambler, M. Elvang-Gøransson, and J. Fox. A logic of argumenta-tion for reasoning under uncertainty. Computational Intelligence, 11:113–131,1995.

Y. Labrou, T. Finin, and Y. Peng. Agent communication languages: The currentlandscape. Intelligent Systems, 14(2):45–52, 1999.

K. Larson and T. Sandholm. An alternating offers bargaining model for com-putationally limited agents. In C. Castelfranchi and L. Johnson, editors,Proceedings of the 1st International Joint Conference on Autonomous Agentsand Multi-Agent Systems (AAMAS-2002), pages 135–142, 2002.

G. L. Lilien, P. Kotler, and S. K. Moorthy. Marketing Models. Prentice-HallPress, USA, 1992.

R. Loui. Defeat among arguments: a system of defeasible inference. Computa-tional Intelligence, 3:100–106, 1987.

43

X. Luo, N. R. Jennings, N. Shadbolt, H.-F. Leung, and J. H.-M. Lee. A fuzzyconstraint-based knowledge model for bilateral, multi-issue negotiations incompetitive environments. Artificial Intelligence, 148(1–2):53–102, 2003.

X. Luo, C. Zhang, and N. R. Jennings. A hybrid model for sharing informationbetween fuzzy, uncertain and default reasoning models in multi-agent sys-tems. International Journal of Uncertainty, Fuzziness and Knowledge-BasedSystems, 10(4):401–450, 2002.

N. Maudet and B. Chaib-draa. Commitment-based and dialogue-game basedprotocols–news trends in agent communication language. Knowledge Engi-neering Review, 17(2):157–179, 2003.

J. Mayfield, Y. Labrou, and T. Finin. Evaluating kqml as an agent communi-cation language. In M. J. Wooldridge, J. P. Muller, and M. Tambe, editors,Intelligent Agents II, volume 1039 of Lecture Notes in Computer Science,pages 347–360. Springer Verlag, Berlin, Germany, 1996.

P. McBurney. Rational Interaction. PhD thesis, University of Liverpool, 2002.

P. McBurney, S. Parsons, and M. Wooldridge. Desiderata for agent argumen-tation protocols. In C. Castelfranchi and L. Johnson, editors, Proceedings ofthe 1st International Joint Conference on Autonomous Agents and MultiagentSystems (AAMAS-2002), pages 402–409, New York, USA, 2002. ACM Press.

P. McBurney, R. M. van Eijk, S. Parsons, and L. Amgoud. A dialogue-gameprotocol for agent purchase negotiations. Journal of Autonomous Agents andMulti-Agent Systems, 7(3):235–273, 2003.

D. L. McGuinness. Ontologies and online commerce. IEEE Intelligent Systems,16(1):8–14, January/February 2001.

D. L. McGuinness and F. van Harmelen. Web ontology language (owl):Overview. Technical report, W3C Working Draft, 31 March 2003.

B. Moulin and B. Chaib-Draa. A review of distributed artificial intelligence. InG. O’Hare and N. R. Jennings, editors, Foundations of Distributed ArtificialIntelligence, chapter 1, pages 3–55. John Wiley & Sons, 1996.

R. Mukherjee, B. Banerjee, and S. Sen. Learning mutual trust. In R. Falcone,M. P. Singh, and Y.-H. Tan, editors, Trust in Cyber-societies, volume 2246 ofLecture Notes in Computer Science, pages 145–158. Springer Verlag, Berlin,Germany, 2001.

M. J. Osborne and A. Rubinstein. A Course in Game Theory. MIT Press,Cambridge MA, USA, 1994.

P. Panzarasa and N. R. Jennings. Social influence, negotiation and cognition.Simulation Modelling Practice and Theory, 10(5–7):417–453, 2002.

44

P. Panzarasa, N. R. Jennings, and T. J. Norman. Formalising collaborativedecision making and practical reasoning in multi-agent systems. Journal ofLogic and Computation, 12(1):55–117, 2002.

S. Parsons and P. Giorgini. An approach to using degrees of belief in bdiagents. In B. Bouchon-Meunier, R. R. Yager, and L. A. Zadeh, editors, In-formation, Uncertainty and Fusion. Kluwer Academic Publishers, Dordrecht,Netherlands, 2000.

S. Parsons, C. Sierra, and N. Jennings. Agents that reason and negotiate byarguing. Journal of Logic and Computation, 8(3):261–292, 1998.

S. Parsons, M. Wooldridge, and L. Amgoud. An analysis of formal inter-agentdialogues. In C. Castelfranchi and L. Johnson, editors, Proceedings of the1st International Joint Conference on Autonomous Agents and MultiAgentSystems (AAMAS-2002), pages 394–401, New York, USA, 2002. ACM Press.

S. Parsons, M. Wooldridge, and L. Amgoud. On the outcomes of formalinter-agent dialogues. In J. Rosenschein, T. Sandholm, M. Wooldridge, andM. Yokoo, editors, Proceedings of the 2nd International Joint Conference onAutonomous Agents and Multiagent Systems (AAMAS-2003), pages 616–623.ACM Press, 2003.

H. V. D. P. Parunak. Industrial and practical applications of dai. In G. Weiss,editor, Multiagent Systems: A Modern Approach to Distributed Artificial In-telligence, pages 377–421. MIT Press, Cambridge MA, USA, 1999.

H. Prakken and G. Sartor. The role of logic in computational models of legalargument: a criticial survey. In A. Kakas and F. Sadri, editors, ComputationalLogic: From Logic Programming into the Future (In honour of Bob Kowalski),volume 2048 of Lecture Notes in Computer Science, pages 342–3. SpringerVerlag, Berlin, Germany, 2001.

H. Prakken and G. Vreeswijk. Logics for defeasible argumentation. In D. Gab-bay and F. Guenthner, editors, Handbook of Philosophical Logic, volume 4,pages 219–318. Kluwer Academic Publishers, Dordrecht, Netherlands, secondedition, 2002.

D. Pynadath and M. Tambe. M. multiagent teamwork: Analyzing key teamworktheories and models. In C. Castelfranchi and L. Johnson, editors, Proceedingsof the 1st International Joint Conference on Autonomous Agents and Multi-agent Systems (AAMAS-2002), pages 873–880, New York, USA, 2002. ACMPress.

I. Rahwan, R. Kowalczyk, and H. H. Pham. Intelligent agents for automatedone-to-many e-commerce negotiation. In M. Oudshoorn, editor, Proceedingsof the 25th Australasian conference on Computer science, pages 197–204. Aus-tralian Computer Society Press, 2002.

45

I. Rahwan, P. McBurney, and L. Sonenberg. Towards a theory of negotiationstrategy (a preliminary report). In S. Parsons and P. Gmytrasiewicz, editors,Proceedings of the 5th Workshop on Game Theoretic and Decision TheoreticAgents (GTDT-2003), pages 73–80, 2003a.

I. Rahwan, L. Sonenberg, and F. Dignum. On interest-based negotiation. InF. Dignum, editor, Advances in Agent Communication, volume 2922 of Lec-ture Notes in Artificial Intelligence. Springer Verlag, Berlin, Germany, 2003b.

I. Rahwan, L. Sonenberg, and F. Dignum. Towards interest-based negotiation.In J. Rosenschein, T. Sandholm, M. Wooldridge, and M. Yokoo, editors,Proceedings of the 2nd International Joint Conference on Autonomous Agentsand Multiagent Systems (AAMAS-2003), pages 773–780. ACM Press, 2003c.

S. Ramchurn, C. Sierra, L. Godo, and N. R. Jennings. A computational trustmodel for multi-agent interactions based on confidence and reputation. InProceedings of the workshop on Deception, Fraud and Trust in Agent Societies,pages 69–75, Melbourne, Australia, 2003a.

S. D. Ramchurn, N. R. Jennings, and C. Sierra. Persuasive negotiation forautonomous agents: a rhetorical approach. In C. Reed, F. Grasso, andG. Carenini, editors, Proceedings of the IJCAI Workshop on ComputationalModels of Natural Argument, pages 9–17. AAAI Press, 2003b.

A. S. Rao and M. P. Georgeff. Formal models and decision procedures formulti-agent systems. Technical Report 61, Australian Artificial IntelligenceInstitute, 171 La Trobe Street, Melbourne, Australia, 1995.

J. A. Rodriguez-Aguılar and C. Sierra. Enabling open agent institutions. InK. Dautenhahn, A. H. Bond, L. Canamero, and B. Edmonds, editors, SociallyIntelligent Agents: Creating relationships with computers and robots, pages259–266. Kluwer Academic Publishers, 2002.

J. Rosenschein and G. Zlotkin. Rules of Encounter: Designing Conventions forAutomated Negotiation among Computers. MIT Press, Cambridge MA, USA,1994.

J. C. Rubiera, J. M. M. Lopez, and J. D. Muro. A fuzzy model of reputation inmulti-agent systems. In Proceedings of the 5th International Conference onAutonomous Agents, pages 25–26. ACM Press, 2001.

A. Rubinstein. Modeling Bounded Rationality. MIT Press, Cambridge MA,USA, 1997.

J. Sabater and C. Sierra. Regret: a reputation model for gregarious societies. InC. Castelfranchi and L. Johnson, editors, Proceedings of the 1st InternationalJoint Conference on Autonomous Agents and Multi-Agent Systems (AAMAS-2002), pages 475–482. ACM Press, 2002.

46

F. Sadri, F. Toni, and P. Torroni. Dialogues for negotiation: agent varietiesand dialogue sequences. In J.-J. Meyer and M. Tambe, editors, IntelligentAgent Series VIII: Proceedings of the 8th International Workshop on AgentTheories, Architectures, and Languages (ATAL 2001), volume 2333 of LectureNotes in Computer Science, pages 69–84. Springer Verlag, Berlin, Germany,2001a.

F. Sadri, F. Toni, and P. Torroni. Logic agents, dialogues and negotiation: anabductive approach. In K. Stathis and M. Schroeder, editors, Proceedings ofthe AISB 2001 Symposium on Information Agents for E-Commerce, 2001b.

F. Sadri, F. Toni, and P. Torroni. Abductive logic programming architecturefor negotiating agents. In Proceedings of the 8th European Conference onLogics in Artificial Intelligence (JELIA-2002), volume 2424 of Lecture Notesin Computer Science, pages 419–431. Springer Verlag, Berlin, Germany, 2002.

L. Samuelson. Evolutionary Games and Equilibrium Selection. MIT Press,Cambridge MA, USA, 1998.

T. Sandholm. Algorithm for optimal winner determination in combinatorialauctions. Artificial Intelligence, 135(1–2):1–54, 2002a.

T. Sandholm. eMediator: A next generation electronic commerce server. Com-putational Intelligence, Special issue on Agent Technology for Electronic Com-merce, 18(4):656–676, 2002b.

T. Sandholm and R. Crites. Multiagent reinforcement learning in the iteratedprisoner’s dilemma. Biosystems, 37:147–166, 1995. Special Issue on the Pris-oner’s Dilemma.

J. Searle. Speech Acts: An Essay in the Philosophy of Language. CambridgeUniversity Press, New York, USA, 1969.

S. Sen, A. Biswas, and S. Debnath. Believing others: Pros and cons. ArtificialIntelligence, 142(2):179–203, 2002.

C. Sierra, P. Faratin, and N. R. Jennings. A service-oriented negotiation modelbetween autonomous agents. In Proceedings of the 8th European Workshopon Modeling Autonomous Agents in a Multi-Agent World (MAAMAW-1997),pages 17–35, Ronneby, Sweden, 1997.

C. Sierra, N. R. Jennings, P. Noriega, and S. Parsons. A framework forargumentation-based negotiation. In M. Singh, A. Rao, and M. Wooldridge,editors, Intelligent Agent IV: 4th International Workshop on Agent Theories,Architectures and Languages (ATAL-1997), volume 1365 of Lecture Notes inArtificial Intelligence, pages 177–192. Springer Verlag, Berlin, Germany, 1998.

M. P. Singh. A social semantics of agent communication lauguages. InF. Dignum and M. Greaves, editors, Issues in Agent Communication, volume1916 of Lecture Notes in Computer Science, pages 31–45. Springer Verlag,Berlin, Germany, 2000.

47

A. M. Spence. Market Signaling: Informational Transfer in Hiring and RelatedScreening Processes. Harvard Economic Studies. Harvard University Press,Cambridge MA, USA, 1974.

K. Sycara. Arguments of persuasion in labour mediation. In Proceedings of the9th International Joint Conference on Artificial Intelligence, pages 294–296,1985.

K. Sycara. The PERSUADER. In D. Shapiro, editor, The Encyclopedia ofArtificial Intelligence. John Wiley & Sons, January 1992.

F. Tohme. Negotiation and defeasible reasons for choice. In Proceedings of theStanford Spring Symposium on Qualitative Preferences in Delbiberation andPractical Reasoning, pages 95–102, 1997.

P. Torroni. A study on the termination of negotiation dialogues. In C. Castel-franchi and L. Johnson, editors, Proceedings of the 1st International JointConference on Autonomous Agents and Multiagent Systems (AAMAS-2002),pages 1223–1230, New York, USA, 2002. ACM Press.

P. Torroni and F. Toni. Extending a logic based one-to-one negotiation frame-work to one-to-many negotiation. In P. P. A. Omicini and R. Tolksdorf,editors, 2nd International Workshop on Engineering Societies in the AgentsWorld II (ESAW 2001), volume 2203 of Lecture Notes in Computer Science,pages 105–118. Springer Verlag, Berlin, Germany, 2001.

D. Traum. Speech acts for dialogue agents. In M. Wooldridge and A. Rao, ed-itors, Foundations And Theories Of Rational Agents, pages 169–201. KluwerAcademic Publishers, Dordrecht, Netherlands, 1999.

H. R. Varian. Mechanism design for computerized agents. In Proceedings of theUSENIX Workshop on Electronic Commerce, New York, July 1995.

J. von Neuman and O. Morgenstern. The Theory of Games and EconomicBehaviour. Princeton University Press, Princeton NJ, USA, 1944.

B. von Stengel. Computing equilibria for two-person games. In R. Aumann andS. Hart, editors, Handbook of game theory, volume 3, chapter 45, pages 1723–1759. Elsevier Science Publishers (North-Holland), Amsterdam, Netherlands,2002.

G. Vreeswijk. Abstract argumentation systems. Artificial Intelligence, 90(1–2):225–279, 1997.

D. N. Walton and E. C. W. Krabbe. Commitment in Dialogue: Basic Conceptsof Interpersonal Reasoning. SUNY Press, Albany NY, USA, 1995.

M. Wooldridge and S. Parsons. Languages for negotiation. In W. Horn, editor,Proceedings of the 14th European Conference on Artificial Intelligence (ECAI-2000), pages 393–397. John Wiley & Sons, 2000.

48

M. J. Wooldridge. Agent-based software engineering. IEE Proceedings on Soft-ware Engineering, 144(1):26–37, 1997.

M. J. Wooldridge. Reasoning about Rational Agents. MIT Press, 2000.

M. J. Wooldridge. An Introduction to MultiAgent Systems. John Wiley & Sons,Chichester, England, 2002.

P. R. Wurman. Market Structure and Multidimensional Auction Design forComputational Economies. PhD thesis, University of Michigan, August 1999.

P. R. Wurman, M. P. Wellman, and W. E. Walsh. A parametrization of theauction design space. Games and Economic Behavior, 35(1–2):304–338, 2001.

M. Yokoo. The distributed constraint satisfaction problem: Formalization andalgorithms. IEEE Transaction On Knowledge and Data engineering, 10(5):673–685, September 1998.

M. A. Young. Rational Games: A Philosophy of Business Negotiation fromPractical Reason. Quorum Books, Westport, CT, USA, 2001.

B. Yu and M. P. Singh. Distributed reputation management for electroniccommerce. Computational Intelligence, 18(4):535–549, 2002.

D. Zeng and K. Sycara. Benefits of learning in negotiation. In Proceedingsof the 14th National Conference on Artificial Intelligence and 9th InnovativeApplications of Artificial Intelligence Conference (AAAI-97/IAAI-97), pages36–42, Menlo Park, USA, 27–31 1997. AAAI Press.

49

argumentation-based negotiation

Documents